Paper deep dive
HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian Splatting
Wei Zhang, Shengkai Yu, Shiqiang Gong, Qi Zhang, Qiang Li, Qi Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/17/2026, 5:11:25 AM
Summary
The paper introduces HiCo-GS, a framework for high-fidelity urban reconstruction using Octree Gaussian Splatting. It addresses two main limitations in existing methods: cross-level feature isolation and lack of geometric supervision. HiCo-GS proposes Cross-Level Context Aggregation (CLCA) to enable bidirectional information flow between octree levels and Depth-Normal Geometric Consistency (DNGC) regularization to enforce geometric coherence. The authors also introduce the China-Pagoda dataset to test reconstruction on complex architectural structures.
Entities (10)
Relation Signals (9)
Wei Zhang → affiliatedwith → Northwestern Polytechnical University
confidence 95% · Wei Zhang ∗ zhangwei707@mail.nwpu.edu.cn School of Computer Science Northwestern Polytechnical University
HiCo-GS → evaluatedon → China-Pagoda
confidence 95% · Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality
HiCo-GS → uses → Cross-Level Context Aggregation
confidence 95% · We present HiCo-GS, a high-fidelity reconstruction framework with two complementary modules. Cross-Level Context Aggregation (CLCA) enables...
HiCo-GS → uses → Depth-Normal Geometric Consistency
confidence 95% · Depth-Normal Geometric Consistency (DNGC) regularization enforces agreement between rendered normals and depth-derived normals...
HiCo-GS → evaluatedon → UrbanScene3D
confidence 90% · Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality
HiCo-GS → evaluatedon → Mill19
confidence 90% · Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality
HiCo-GS → evaluatedon → MatrixCity
confidence 90% · Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Octree-based anchor Gaussian Splatting has emerged as a scalable representation for city-scale novel view synthesis, where multi-level anchors adaptively capture scene content from coarse building structures to fine architectural details. However, we identify a fundamental limitation in existing methods: cross-level feature isolation, where each level's anchor features are optimized independently with no inter-level communication, causing color drift on building facades and over-smoothing in textured regions. We present HiCo-GS, a high-fidelity reconstruction framework with two complementary modules. Cross-Level Context Aggregation (CLCA) enables bidirectional hierarchical prior injection by leveraging the octree's spatial containment structure to aggregate per-level context vectors into parent-self-child triplets, fused via a lightweight MLP with residual connection. Coarse-level structural priors flow down to inform fine-level anchors, while fine-level detail statistics feed back to prevent over-smoothing, at negligible computational overhead. Depth-Normal Geometric Consistency (DNGC) regularization enforces agreement between rendered normals and depth-derived normals through an alpha-weighted consistency loss, complemented by edge-aware smoothness losses with progressive warmup that exploit the strong planar priors ubiquitous in urban geometry to suppress floating artifacts. We further introduce the China-Pagoda dataset comprising 8 ancient Chinese pagodas with over 1,200 images each, featuring dense ornamental carvings, curved multi-layer eaves, and repetitive fine-grained textures. Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate that HiCo-GS achieves state-of-the-art rendering quality and substantially cleaner geometry across real-world and synthetic urban this http URL: this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.14136v1
- Canonical: https://arxiv.org/abs/2608.14136v1
Trouble viewing inline? Open PDF directly →
Full Text
73,158 characters extracted from source content.
Expand or collapse full text
HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian Splatting Wei Zhang ∗ zhangwei707@mail.nwpu.edu.cn School of Computer Science Northwestern Polytechnical University Xi’an, China Shengkai Yu ∗ yyusober@mail.nwpu.edu.cn Northwestern Polytechnical University Xi’an, China Shiqiang Gong ∗ gongshiqiang@mail.nwpu.edu.cn Northwestern Polytechnical University Xi’an, China Qi Zhang nwpuqzhang@gmail.com vivo BlueImage Lab vivo Mobile Communication Co., Ltd. Hangzhou, China Qiang Li qiangli@nwpu.edu.cn Northwestern Polytechnical University Xi’an, China Qi Wang † crabwq@gmail.com Northwestern Polytechnical University Xi’an, China Building Our Mesh / Texture Our Mesh / Texture Residence (a) Hico-GS (Ours)(b) Citygs-X depth Lingshan normal BlockGaussian Citygs-v1 Mega-NeRF Switch-NeRF Momentum-GS CityGS-X HiCo-GS(Ours) Yunjusi Floating Artifacts. Duobaota Detail Lost. Figure 1: HiCo-GS achieves high-fidelity urban reconstruction with cleaner geometry. Left: Mesh, rendered RGB, depth, and normal comparisons on Mill19 and UrbanScene3D between (a) HiCo-GS and (b) CityGS-X. Right: Representative scenes from our China-Pagoda dataset featuring extreme architectural complexity. Bottom: Radar charts summarizing quantitative results across both benchmarks. HiCo-GS (red) consistently outperforms prior methods. Abstract Octree-based anchor Gaussian Splatting has emerged as a scalable representation for city-scale novel view synthesis, where multi- level anchors adaptively capture scene content from coarse building ∗ These authors contributed equally to this work. † Corresponding author. This work is licensed under a Creative Commons Attribution-NonCommercial- NoDerivatives 4.0 International License. M ’26, Rio de Janeiro, Brazil © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2213-4/2026/11 https://doi.org/10.1145/3767308.3836546 structures to fine architectural details. However, we identify a funda- mental limitation in existing methods: cross-level feature isolation, where each level’s anchor features are optimized independently with no inter-level communication, causing color drift on building facades and over-smoothing in textured regions. We present HiCo- GS, a high-fidelity reconstruction framework with two complemen- tary modules. Cross-Level Context Aggregation (CLCA) enables bidirectional hierarchical prior injection by leveraging the octree’s spatial containment structure to aggregate per-level context vectors into parent-self-child triplets, fused via a lightweight MLP with residual connection. Coarse-level structural priors flow down to inform fine-level anchors, while fine-level detail statistics feed back to prevent over-smoothing, at negligible computational overhead. arXiv:2608.14136v1 [cs.CV] 14 Aug 2026 M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Depth-Normal Geometric Consistency (DNGC) regularization en- forces agreement between rendered normals and depth-derived nor- mals through an alpha-weighted consistency loss, complemented by edge-aware smoothness losses with progressive warmup that exploit the strong planar priors ubiquitous in urban geometry to suppress floating artifacts. We further introduce the China-Pagoda dataset comprising 8 ancient Chinese pagodas with over 1,200 im- ages each, featuring dense ornamental carvings, curved multi-layer eaves, and repetitive fine-grained textures. Extensive experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demon- strate that HiCo-GS achieves state-of-the-art rendering quality and substantially cleaner geometry across real-world and synthetic urban benchmarks. Code: https://github.com/WZ-CS/HiCo-GS. CCS Concepts • Computing methodologies→ Reconstruction. Keywords 3D Gaussian Splatting, Large-Scale Urban Reconstruction, Octree Level-of-Detail, Geometric Regularization ACM Reference Format: Wei Zhang, Shengkai Yu, Shiqiang Gong, Qi Zhang, Qiang Li, and Qi Wang. 2026. HiCo-GS: Hierarchical Context Aggregation and Geomet- ric Consistency for Octree Gaussian Splatting. In Proceedings of the 34th ACM International Conference on Multimedia (M ’26), November 10–14, 2026, Rio de Janeiro, Brazil. ACM, New York, NY, USA, 21 pages. https: //doi.org/10.1145/3767308.3836546 1 Introduction 3D Gaussian Splatting (3DGS) [12,39] has established a new par- adigm for real-time novel view synthesis by representing scenes as collections of explicit Gaussian primitives rendered through tile-based differentiable rasterization. To scale this representation to city-level environments [42,43], recent works [8,20,22,25] adopt octree-based anchor structures with level-of-detail (LOD) control, where multi-level anchors reside at different voxel resolu- tions. Coarse-level anchors capture building-scale geometry, while fine-level anchors encode local architectural details such as win- dow frames and surface textures. Neural Gaussian generation then decodes each anchor’s learned feature into a local cluster of Gaus- sians conditioned on the viewing direction, enabling compact yet expressive scene representation. These methods have demonstrated strong scalability through distributed training strategies, making large-scale urban reconstruction increasingly practical. Despite their success, we identify two complementary limitations in existing octree-anchor methods. The first is cross-level feature isolation: as illustrated in Fig. 2 (top), each level’s anchor features are optimized independently through backpropagation with no inter-level communication. A fine-level anchor encoding a window frame has no access to the coarse-level context of the wall it be- longs to, and conversely, a coarse-level anchor is unaware of the local detail density beneath it. This isolation leads to color drift on building facades and over-smoothing in richly textured regions. The second limitation is the absence of geometric supervision: the rasterization stage produces two independent normal estimates, one from Gaussian covariance and one from the rendered depth map, yet existing methods impose no constraint requiring these MLP L1 Feat (Voxel S/2 3 ) Level 1 anchor L2 Feat (Voxel S/N 3 ) Level N anchor (a) Existing: Feature Isolation. 3D Gaussian Splatting MLP MLP L0 Feat (Voxel S) Level 0 anchor (b) Ours: Cross-level Context. MLP L1 Feat (Voxel S/2 3 ) Self L2 Feat (Voxel S/N 3 ) Fine / Child MLP MLP L0 Feat (Voxel S) Coarse / Parent C 0 C 1 C 2 R e s i d u a l C o n t e x t M L P L0 L1 L1 L2 L3 L4 L0 L1 L2 L3 L4 L0 Feature embedding (t-SNE) Feature embedding (t-SNE) Avg. cosine dist. ≈ 0.97. Avg. cosine dist. ≈ 0.82 (↓15.5%). Figure 2: Cross-level feature isolation and our solution. Top: In existing octree-based anchor Gaussian Splatting, multi- level anchors (L0 to LN) are optimized independently—each level’s features are decoded by shared MLPs with no inter- level communication. The t-SNE visualization (right) con- firms this: features from all levels form a structureless mix- ture with inter-level cosine distances approaching 0.97, indi- cating near-orthogonal representations. Bottom: Our CLCA enables bidirectional hierarchical prior injection. Each an- chor receives a context triplet from its parent (coarse), self, and child (fine) levels, fused through a lightweight MLP with residual connection. Coarse-level structural priors flow down while fine-level detail statistics feed back. After CLCA (right), features exhibit level-aware clustering with adjacent-level cosine distance reduced by up to 43%. signals to agree. In urban scenes dominated by large planar surfaces, this lack of geometric consistency results in noisy normals, depth discontinuities, and floating artifacts. We present HiCo-GS, a high-fidelity reconstruction framework that addresses both limitations. To resolve cross-level feature iso- lation, we propose Cross-Level Context Aggregation (CLCA), which leverages the octree’s inherent spatial containment structure to enable bidirectional context flow across levels. For each anchor, CLCA looks up the specific parent-level context vector at the coarser voxel that spatially contains it, and aggregates child-level features from finer voxels that fall within its own grid cell. The resulting parent-self-child context triplet is fused through a lightweight MLP with residual connection, allowing coarse-level structural priors to inform fine-level anchors while fine-level detail statistics feed back to coarse-level anchors (Fig. 2, bottom). To address the lack of geometric supervision, we introduce Depth-Normal Geometric Consistency (DNGC) regularization, which enforces agreement between rendered normals and depth-derived normals through an alpha-weighted consistency loss, complemented by edge-aware smoothness terms that encourage planarity on smooth surfaces while preserving sharp boundaries. A progressive warmup schedule ensures that geometric constraints do not interfere with early pho- tometric convergence. Furthermore, to stress-test reconstruction fidelity on geometrically extreme architectures beyond standard urban benchmarks, we introduce the China-Pagoda dataset com- prising 8 ancient Chinese pagodas with over 1,200 images each, featuring dense ornamental carvings, curved multi-layer eaves, and repetitive fine-grained textures. Our contributions are as follows: HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil •We identify the cross-level feature isolation problem in octree- based Gaussian Splatting and propose CLCA, a spatially aware context aggregation module that enables bidirectional hierarchical feature communication across octree levels by exploiting the octree’s spatial containment structure, enrich- ing fine-level features with structural priors and coarse-level features with detail statistics. •We introduce DNGC, a depth-normal geometric consistency regularization with edge-aware smoothness and progres- sive warmup that leverages urban planar priors to suppress floating artifacts and produce cleaner surface geometry. •We construct the China-Pagoda benchmark for evaluating reconstruction under extreme geometric complexity, and demonstrate through extensive experiments on Mill19, Ur- banScene3D, MatrixCity, and China-Pagoda that HiCo-GS achieves state-of-the-art rendering quality and substantially cleaner geometry across real-world and synthetic urban benchmarks. 2 Related Work 3D Gaussian Splatting and Large-Scale Extensions. 3DGS [12] represents scenes as anisotropic Gaussian primitives rendered via differentiable rasterization, with subsequent works improving com- pactness [5,7,23,37,40], anti-aliasing [34,35], and appearance modeling [4,14,41]. Scaffold-GS [22] introduces anchor-based neu- ral Gaussians, which our work builds upon. Scaling to city-level scenes has been addressed via spatial partitioning [2,18,20,33], hi- erarchical LOD [13,25–27,44], and distributed training [8]. Notably, Octree-GS [25] organizes anchors into a multi-level octree with LOD-aware densification, and Hierarchy-GS [13] enables smooth transitions via chunk-based consolidation. Geometric Regularization for Gaussian Splatting. Several works incorporate geometric priors to improve surface quality: SuGaR [9] aligns Gaussians to surfaces for mesh extraction, GOF [36] de- rives opacity fields for surface extraction, and normal-based su- pervision has been explored via rendered consistency [11,29,38] and mono-depth priors [3]. For urban scenes, CityGaussianV2 [21] and ULSR-GS [17] enforce geometric consistency across partitions. Our DNGC enforces consistency between rendered normals and depth-derived normals already produced by the rasterizer, using edge-aware weights that respect urban planar structure. 3 Method 3.1 Preliminaries Anchor-based Octree Gaussian Splatting. Anchor-based meth- ods [22,25] place anchors at octree voxel centers, each storing a learnable feature f 푖 ∈ R 푑 , position p 푖 , scaling s 푖 ∈ R 6 , rotation q 푖 ∈ R 4 , opacity표 푖 , and퐾offset vectors휹 푘 푖 퐾 푘=1 . Each anchor generates퐾neural Gaussians centered at x 푘 푖 =p 푖 +휹 푘 푖 ⊙s [1:3] 푖 , whose color, opacity, and covariance are predicted by shared MLPs conditioned on f 푖 and the viewing direction. Given a camera at position c, each anchor’s predicted level is ˆ 푙 푖 = log 2 (푑 std /∥p 푖 − c∥) log 2 푏 +휖 푖 ,(1) where푑 std is a standard distance from training cameras and휖 푖 a learnable adjustment. An anchor at level 푙 푖 is visible only if 푙 푖 ≤ ˆ 푙 푖 . Rasterization Outputs. Differentiable rasterization produces the RGB image I, accumulated opacity휶, and two independent normal estimates. The rendered normal n 푟 is obtained by alpha-compositing per-Gaussian normals defined as the thinnest axis of each Gauss- ian ellipsoid. The depth-derived normal n 푑 is computed by back- projecting the rendered depth map D and taking finite-difference cross products. We also render a plane distance map d, representing the projection of each point onto its local normal direction. These quantities are used by our DNGC regularization (Sec. 3.4). 3.2 Overview of HiCo-GS Given a set of posed images and a sparse SfM point cloud, our pipeline proceeds as follows (Fig. 3). We first construct a multi-level octree from the point cloud, populating each voxel center with an anchor and its associated attributes (Sec. 3.1). For each training camera, we perform LOD-based visibility filtering via Eq. 1 followed by view-frustum culling to obtain the set of visible anchors. The visible anchors’ featuresf 푖 are then enhanced by our Cross-Level Context Aggregation (CLCA) module (Sec. 3.3), which injects hierarchical context from parent and child octree levels into each anchor’s feature to produce ˆ f 푖 . The enhanced fea- tures, concatenated with the viewing direction and optionally a per-camera appearance embedding, are fed into three shared MLPs to predict per-offset opacity, covariance (scale and rotation), and color. Offsets with positive opacity are retained and combined with anchor positions as x 푘 푖 = p 푖 +휹 푘 푖 ⊙ s [1:3] 푖 to produce the final set of neural Gaussians. Differentiable rasterization then produces the RGB image I, ren- dered normal n 푟 , depth map D (and its derived normal n 푑 ), plane distance d, and opacity map휶. The training loss combines photo- metric reconstruction with our Depth-Normal Geometric Con- sistency (DNGC) regularization (Sec. 3.4): L=L rgb +푤(푡) ( 휆 nc L nc + 휆 ns L ns + 휆 ds L ds ) ,(2) whereL rgb is the standard combination ofℓ 1 and D-SSIM losses, and 푤(푡) is a linear warmup coefficient detailed in Sec. 3.4. 3.3 Cross-Level Context Aggregation In existing octree-anchor methods, each level’s anchor features are optimized independently through backpropagation. A fine-level anchor encoding a window frame has no knowledge of the coarse- level wall it belongs to, and a coarse-level anchor is unaware of the local texture complexity beneath it. We address this cross-level feature isolation through a module that enables bidirectional con- text flow across octree levels. Unlike multi-scale aggregation in point cloud networks [24] that extract hierarchical features from scratch in a single forward pass, CLCA injects cross-level priors into already-learned features during iterative rendering optimization. And unlike UNet-style skip connections across sequential encoder- decoder stages, our octree levels coexist simultaneously under LOD selection, requiring spatially-indexed lookup. Spatial-Aware Level Context. For each octree level푙, we compute spatially-indexed context features. Given voxel size푣 푙 at level푙, each visible anchor is mapped to its grid coordinate g 푙 푖 =⌊( p 푖 −o)/푣 푙 ⌉, M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Sparse Point Clouds Octree / Multi-level Anchors L0 position ∈ℝ 퐀 f enhanced Neural Gaussian Generation Sec. 3.3 Cross-Level Context Aggregation (CLCA) Per-level Anchor Feats L i-1 L i L i+1 L e v e l - w i s e F e a t u r e A g g r e g a t i o n n r from con. n d from depth L1 diff I → |∇I| L ns (normal) L ds (dist) Sec. 3.4 Depth-Normal Geometric Consistency (DNGC) L nc feature ∈ℝ 퐀䐀 Attributes scale ∈ℝ 퐀 offsets ∈ℝ 퐀×퐀 rotation ∈ℝ 퐀 L1 level ∈ℝ 퐀 E n h a n c e d f e a t （x, y, z） 1 / 2 .. MLP cov / MLP opacity MLP color RGB image I Differentiable Rasterization Depth D + dist d Depth → n d Normal n r + alpha alpha e -β|∇I| L rgb (L1+SSIM) C L C A MLPs D N G C l o s s e s Sec. 3.2 Overview of HiCo-GS n r →|∇n_r| d → |∇d| w L = L rgb + w(t) (λ nc L nc + λ ns L ns + λ ds L ds ) w(t) = min( (t - t 0 ) / T, 1 ) (linear warmup). ：Pose / Id / View dir. e -β|∇I| ： Edge Weight. Feat Concat. c p a r e n t MLP context + L0 / L1 ：LN Feat (Voxel S/N). dist d ：distance. alpha：Accumulated opacity. + ：Concat. Structural prior Detail statistics Bidirectional Hierarchical Prior Injection c s e l f c c h i l d ：Gaussian covariance. con. Figure 3: Overview of HiCo-GS. (a) A sparse point cloud is organized into a multi-level octree, where each anchor stores a feature vector, position, level, offsets, scale, and rotation. Visible anchors are enhanced by CLCA to produce ˆ f, decoded by three shared MLPs (conditioned on viewing direction and appearance) into neural Gaussians. Differentiable rasterization produces RGB image I, rendered normal n 푟 , depth D (with derived normal n 푑 ), distance d, and opacity휶. (b) CLCA performs spatially-aware level-wise aggregation to compute per-level context vectors, then assembles a parent-self-child triplet for each anchor via hierarchical context lookup. The triplet is fused byMLP context with a residual connection. (c) DNGC enforces consistency between n 푟 (from Gaussian covariance) and n 푑 (from depth) viaL nc , and applies edge-aware smoothness lossesL ns , L ds weighted by 푒 −훽|∇I| to enforce planarity while preserving sharp boundaries. L0L1L2L3L4L5L6 L0 L1 L2 L3 L4 L5 L6 0.000.56-0.071.00-0.020.470.32 0.560.000.570.510.590.270.41 -0.070.570.000.060.260.490.61 1.000.510.060.000.100.460.79 -0.020.590.260.100.000.250.76 0.470.270.490.460.250.000.87 0.320.410.610.790.760.870.00 (a) Similarity Building L0L1L2L3L4 L0 L1 L2 L3 L4 0.000.140.230.770.28 0.140.000.100.650.07 0.230.100.000.770.34 0.770.650.770.000.73 0.280.070.340.730.00 (b) Similarity Residence L01L12L23L34L45L56 0.00 0.25 0.50 0.75 1.00 Cosine dist. 15% 14% 16% 22% 11% 88% (c) Cosine dist. Building BeforeAfter L01L12L23L34 0.00 0.25 0.50 0.75 1.00 Cosine dist. 31% 41% 64% 29% (d) Cosine dist. Residence BeforeAfter Figure 4: Quantitative analysis of CLCA on Building and Residence. (a, b) Change in inter-level cosine similarity after CLCA (warmer = larger increase). Both scenes show strong positive gains, particularly between fine levels (Building L5–L6:Δ=0.87; Residence L2–L3:Δ=0.77), confirming effective cross-level context flow. (c, d) Adjacent-level cosine distance before and after CLCA. On Building, all pairs show 11–88% reduction, with the finest pair (L5–L6) dropping from 0.97 to 0.12. On Residence, reductions range from 29% to 64% across all level pairs. More details are provided in the supplementary material. where o is the octree origin. Anchors at the same level sharing the same grid coordinate are aggregated via mean pooling: c 푙 (g)= MeanPool f 푖 푖 ∈V 푙 , g 푙 푖 = g ,(3) whereV 푙 is the set of visible anchors at level 푙 . Hierarchical Context Lookup. The octree’s spatial containment provides a natural parent-child relationship: an anchor at level 푙resides within a specific voxel at level푙−1. We compute each anchor’s parent-level grid coordinate as g 푙−1 푖 =⌊( p 푖 −o)/푣 푙−1 ⌉and look up the corresponding context: c parent (a 푖 )= c 푙 푖 −1 (g 푙 푖 −1 푖 ).(4) For child context, we aggregate all level-(푙+1)anchors mapping into the same level-푙 grid cell as anchor a 푖 : c child (a 푖 )= MeanPool n f 푗 푗 ∈V 푙 푖 +1 , g 푙 푖 푗 = g 푙 푖 푖 o ,(5) i.e. features of finer-level anchors whose parent-level grid coordi- nate matches that of a 푖 . For anchors at the coarsest (or finest) active level, the self-level context is used as a fallback for the missing parent (or child). We limit scope to a one-hop triplet (parent-self- child): with푏=2, adjacent levels span a 4×resolution ratio, and more distant information propagates implicitly since each level’s context is itself enriched by its neighbors. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil The parent context c parent carries structural priors—coarse-level information about wall orientation, facade material, and building- scale color that informs fine-level anchors. The child context c child provides detail statistics—fine-level information about local texture density and micro-geometry that prevents coarse-level anchors from over-smoothing. Feature Enhancement. For each visible anchor a 푖 at level푙 푖 with self-context c self =c 푙 푖 (g 푙 푖 푖 ), we assemble the triplet and produce the enhanced feature via a lightweight MLP with residual connection: ˆ f 푖 = f 푖 + MLP 휃 [c parent ; c self ; c child ] ,(6) where[·;·;·]denotes concatenation andMLP 휃 consists of two lin- ear layers with ReLU activation (3푑 → 푑 → 푑, with푑=32). The residual connection ensures that the module acts as a refinement: at initialization, the MLP outputs near-zero values and the rep- resentation reverts to the baseline behavior. The spatial lookup uses hash-based indexing with푂(푁 log푀)complexity, and the MLP adds only 3,168 parameters—negligible relative to the main prediction MLPs. Empirical validation in Fig. 4. 3.4 Depth-Normal Geometric Consistency The rasterization stage produces two independent normal estimates n 푟 and n 푑 (Sec. 3.1) that are derived from fundamentally differ- ent geometric information—Gaussian covariance and depth map, respectively. In existing methods, these two signals are not con- strained to agree, allowing the Gaussian orientations to diverge from the actual surface geometry without penalty. This is particu- larly problematic in urban scenes where large planar regions (e.g. building facades) should exhibit spatially coherent normals. We introduce three complementary regularization terms that leverage this geometric redundancy. Normal Consistency Loss. We enforce agreement between the two normal estimates, weighted by the accumulated opacity to focus on regions with sufficient Gaussian coverage: L nc = Í p 휶(p)· ∥ ˆ n 푟 (p)− ˆ n 푑 (p) ∥ 1 Í p 휶(p) ,(7) where ˆ n 푟 and ˆ n 푑 areℓ 2 -normalized, and the summation is over all pixels p. Edge-Aware Smoothness. Urban scenes exhibit a useful structural property: surface normals and depth should vary smoothly within planar regions (e.g. a wall), while being allowed to change abruptly at object boundaries (e.g. a window frame). We encode this prior through an edge-aware weighting derived from the RGB image gradient: 푤 푥 (p)= exp ( −훽 | ∇ 푥 I(p) |) , 푤 푦 (p)= exp −훽 ∇ 푦 I(p) ,(8) where∇ 푥 and∇ 푦 are horizontal and vertical finite differences of the rendered RGB image averaged across color channels, and훽=10 controls edge sensitivity. This weight approaches 1 in flat regions (encouraging smoothness) and decays toward 0 at color edges (per- mitting discontinuities). Using this weight, we define the normal smoothness and distance smoothness losses: L ns =푤 푥 ·|∇ 푥 n 푟 |+푤 푦 ·|∇ 푦 n 푟 |,(9) L ds = 푤 푥 ·|∇ 푥 d|+푤 푦 ·|∇ 푦 d|,(10) Figure 5: Overview of the China-Pagoda dataset. More details are provided in the supplementary material. where(·)denotes the spatial mean. Both losses share the same edge weight푤from the RGB image:L ns penalizes normal discontinuities on smooth surfaces (e.g. a wall with an inconsistently oriented Gaussian), whileL ds penalizes distance jumps that indicate floating artifacts. Progressive Warmup. Geometric regularization requires reason- ably converged depth and normal estimates; applying it too early when these signals are noisy can hinder RGB convergence. We therefore use a linear warmup: 푤(푡)= min 푡 −푡 0 푇 , 1 ,(11) where푡 0 is the iteration at which geometry rendering is activated and푇controls the ramp-up duration. This allows the photometric loss to dominate early training while geometric constraints are gradually introduced. We use휆 nc =0.05,휆 ns =0.01,휆 ds =0.01, and 푇= 10,000 across all experiments. 3.5 China-Pagoda Benchmark Existing urban reconstruction benchmarks such as Mill19 [28] and UrbanScene3D [19] are dominated by modern buildings with regu- lar planar facades. Methods evaluated on these scenes can achieve high scores primarily by modeling flat surfaces well, while weak- nesses in handling complex non-planar geometry remain unex- posed. To provide a complementary evaluation axis, we introduce the China-Pagoda dataset, a collection of 8 ancient Chinese pago- das whose architectural characteristics differ fundamentally from modern urban structures. Data Collection. We capture 8 pagodas across northern China, spanning diverse styles from the Tang, Liao, and Qing dynasties (Fig. 5). The structures range from 7-story brick pagodas with dense surface carvings to 13-story glazed-tile towers with elab- orate bracket systems. For each pagoda, we collect over 1,200 raw images using drone-based aerial photography and ground-level captures, totaling over 10,000 images. Multi-altitude drone passes ensure complete coverage from base to spire, while ground-level captures provide close-range views of surface details. Geometric Challenges. The China-Pagoda scenes exhibit three characteristics that stress-test reconstruction methods: dense orna- mental carvings with millimeter-level relief detail that push the res- olution limits of anchor-based representations, curved multi-layer eaves with complex bracket systems (dougong) whose non-planar, self-occluding geometry cannot be captured by simple planar as- sumptions, and repetitive fine-grained brick and tile textures that create strong visual ambiguity across viewpoints. The combina- tion of intricate multi-scale geometry, non-planar self-occluding structures, and pervasive texture repetition exposes limitations that standard planar-dominated scenes fail to reveal. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Table 1: Quantitative comparison on Mill19 [28] and UrbanScene3D [19].↑: higher is better,↓: lower is better. Best and second best results are highlighted.†: without decoupled appearance encoding. Method BuildingRubbleResidenceSci-Art SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓ Without geometric optimization Mega-NeRF [28]0.54720.920.4540.55324.060.5080.62822.080.4010.77025.600.312 Switch-NeRF [19]0.57921.540.3970.56224.310.4780.65422.570.3520.79526.510.271 VastGaussian† [18]0.72821.800.2250.74225.200.2640.69921.010.2610.76122.640.261 3DGS [12]0.73822.530.2140.72525.510.3160.74522.360.2470.79124.130.262 DoGaussian [2] 0.75922.730.2040.76525.780.2570.74021.940.2440.80424.420.219 Momentum-GS [6]0.815 23.230.1940.82725.930.2010.81822.210.1970.85623.020.205 CityGaussian [20]0.77821.550.2460.81325.770.2280.81322.000.2110.83721.390.230 With geometric optimization SuGaR [9]0.50717.760.4550.57720.690.4530.60318.740.4060.69818.600.349 NeuS [30]0.46318.010.6110.48020.460.6180.50317.850.5330.63318.620.472 Neuralangelo [16]0.58217.890.3220.62520.180.3140.64418.030.2630.76919.100.231 PGSR [1]0.48016.120.5730.72823.090.3340.74620.570.2890.79919.720.275 PGSR+VastGS 0.72021.630.3000.76825.320.274– CityGaussianV2 [21]0.65019.070.3970.72023.750.3220.76921.150.2340.81020.660.266 CityGS-X [8]0.81722.76 0.1910.82326.150.2100.81922.440.1940.86722.770.179 HiCo-GS (Ours)0.78922.670.2310.863 28.09 0.1920.844 24.41 0.1660.894 26.64 0.163 Table 2: Comparison with existing 3D reconstruction bench- marks. China-Pagoda provides dense per-scene coverage of geometrically extreme structures, complementing existing urban and cultural heritage datasets. DatasetScenesTotal imgsImgs/sceneTypeGeometry Mill19 [28]2 ∼4K ∼2KUrbanPlanar facades UrbanScene3D [19]6 ∼5K ∼0.8KUrbanPlanar facades MatrixCity [15]2 ∼10K ∼5KSyntheticMixed China-Pagoda (Ours)8 ∼10K+ ∼1.2K+HeritageExtreme Table 2 compares China-Pagoda with existing benchmarks. While not the largest in total image count, it is uniquely positioned in per-scene density and geometric complexity. For each scene, we provide raw images, COLMAP-reconstructed camera poses and sparse point clouds, and a standard train/test split. The dataset will be publicly released. 4 Experiments 4.1 Implementation Details We implement HiCo-GS on top of the CityGS-X[8] codebase. All experiments are conducted on 4×NVIDIA RTX 4090 GPUs with distributed training. We train for 30,000 iterations with a batch size equal to the number of GPUs. The CLCA module uses a two- layer MLP (96→32→32) with ReLU activation, sharing the same learning rate schedule as the color MLP. For DNGC, we set 휆 nc =0.05,휆 ns =0.01,휆 ds =0.01, edge sensitivity훽=10, and warmup duration푇=10,000 iterations. The geometry rendering is activated at iteration푡 0 following the same schedule as CityGS-X. We use the default octree configuration with branching factor푏=2, feature dimension푑=32, and퐾=5 offsets per anchor. All other hyperparameters follow CityGS-X. 4.2 Main Results Novel View Synthesis. In Tab. 1 and Fig. 6, we conduct both quantitative and qualitative comparisons on Mill19 [28] and Urban- Scene3D [19] to evaluate the rendering quality of recent large-scale reconstruction methods w/ and w/o geometric optimizations. It is evident that HiCo-GS achieves state-of-the-art performance across all four scenes and all metrics, outperforming both categories of methods by a significant margin. Compared to our direct baseline CityGS-X, we observe substantial improvements, e.g., a +1.94 dB PSNR gain and a 0.018 LPIPS reduction on Rubble, and a +3.87 dB PSNR gain with a 0.016 LPIPS reduction on Sci-Art. Notably, HiCo- GS also surpasses the best methods without geometric optimization: compared to Momentum-GS [6], we achieve a +2.16 dB PSNR im- provement on Rubble and a +2.20 dB improvement on Residence. These improvements are most pronounced on scenes with complex multi-scale structures, where cross-level context communication and geometric regularization are most beneficial. Surface Reconstruction. In Tab. 4, we compare our method with other surface reconstruction methods on MatrixCity [15]. The ex- perimental results demonstrate that HiCo-GS achieves state-of-the- art performance with a PSNR of 27.93 dB and an F1 score of 0.599, surpassing CityGS-Xby 0.35 dB in PSNR and 0.018 in F1. The gains are more moderate than on real-world datasets, which is expected since MatrixCity’s synthetic rendering produces perfectly consis- tent appearances across views, reducing the impact of cross-level context aggregation on appearance modeling. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil H i c o - G S : r e n d e r e d R G B \ n o r m a l C i t y G S - x : r e n d e r e d R G B \ n o r m a l Residence 001720 Residence 001949 Residence 002462 Residence 000370 Residence 001723 Figure 6: Qualitative results of our method and other methods in large-scale reconstruction datasets UrbanScene. Figure 7: Qualitative results of ours and other methods in image and depth rendering on Mill-19 and Urbanscene3D datasets. Fig. 7 presents a comparison between our method and other reconstruction approaches on UrbanScene3D. Our method yields notably smoother normals on planar wall surfaces while preserving sharp transitions at structural edges, validating the effectiveness of DNGC regularization. In contrast, baseline methods produce noisy and inconsistent normal maps, particularly in large planar regions such as building facades and rooftops. Additionally, Fig. 8 presents qualitative mesh and texture comparisons between our method and CityGS-Xon the Residence scene. It can be observed that HiCo-GS produces more detailed and cleaner surface structures with fewer floating artifacts, closely resembling the actual geometry of the scene. In contrast, CityGS-Xexhibits fragmented surfaces and geometric inaccuracies, particularly on building facades. Evaluation on China-Pagoda. In Tab. 3, we evaluate HiCo-GS on our proposed China-Pagoda benchmark to assess performance under extreme geometric complexity. HiCo-GS outperforms all baselines across the majority of scenes and metrics, with improve- ments more substantial than on standard urban benchmarks. For example, on Yuhuangta we achieve a +6.62 dB PSNR gain and a 0.105 LPIPS reduction over CityGS-X, and on Lingshan we reach 32.95 dB PSNR with an SSIM of 0.965. On Baoanta, HiCo-GS improves upon the second-best method SplatCo [32] by +2.19 dB PSNR. Across all eight scenes, HiCo-GS achieves 22 out of 24 first-place rankings, with the remaining two on Beita (SSIM and LPIPS) narrowly taken by CityGS-X (0.770 vs. 0.770 and 0.350 vs. 0.357). Ablation Analysis. As shown in Table 5, both CLCA and DNGC independently yield substantial improvements over the baseline, with average PSNR gains of +3.20 dB and +2.78 dB respectively, confirming that the two modules address complementary failure modes. CLCA shows particularly strong gains on Rubble (+5.72 dB) M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Table 3: Results on China-Pagoda benchmark.↑: higher is better,↓: lower is better. Method BaoantaBeitaDuobaofotaLingshan SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓ CityGS-v1 [20]0.74522.790.3690.75923.920.4270.80623.520.3080.87925.310.211 CityGS-X [8]0.89126.330.1680.77023.17 0.3500.81122.490.2170.95028.770.081 SplatCo [32]0.82727.050.2730.64321.290.5780.79225.880.3040.55016.410.513 BlockGaussian [31] 0.76023.480.3490.68621.740.4590.70020.290.4150.81625.060.269 Momentum-GS [6]0.81523.790.3590.71422.980.4620.79723.900.2570.92127.240.160 HiCo-GS (Ours)0.928 29.51 0.1330.770 24.010.3570.886 28.56 0.1600.965 32.95 0.066 Method YingsiYongwangtaYuhuangtaYunjusi SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓SSIM↑ PSNR↑ LPIPS↓ CityGS-v1 [20]0.67421.540.4480.73722.380.3800.78322.910.3240.80423.500.299 CityGS-X [8]0.61217.220.4020.82421.970.2150.86524.030.1600.89225.22 0.129 SplatCo [32]0.64023.060.5250.77125.620.3530.85527.960.2380.80123.880.273 BlockGaussian [31]0.59519.440.4790.63920.800.4580.79824.180.3140.70420.120.382 Momentum-GS [6]0.61820.500.4840.80521.900.3010.74322.030.3520.82924.180.227 HiCo-GS (Ours)0.813 24.82 0.2640.866 26.68 0.1950.926 30.65 0.1100.903 26.820.132 Residence Hico-GS PSNR:25.42 Residence CityGS-X PSNR:22.44 Figure 8: Qualitative mesh and texture comparison between CityGS-X and our method on the Residence. Table 4: Results on MatrixCity [15]. Best and second best results are highlighted. MethodPSNR↑P↑R↑F1↑ NeuS [30]16.76FAILFAILFAIL Neuralangelo [16]19.220.0800.0830.081 SuGaR [9]OOMOOMOOMOOM GOF [36]17.42FAILFAILFAIL 2DGS [10]21.350.2070.3900.270 CityGS [20]27.460.3620.6370.462 CityGS-V2 [21]27.230.4410.7520.556 CityGS-X [8]27.580.4440.8400.581 HiCo-GS (Ours)27.930.4610.8560.599 where multi-scale structural context is critical. DNGC delivers the largest single-scene improvement on Rubble (+5.18 dB), reflecting its effectiveness in producing geometrically coherent surfaces. The Table 5: Ablation study. PSNR (dB)↑on Mill19/UrbanScene3D. CLCADNGCBuild.Rub.Res.Sci.Avg. Time –22.4020.5721.8122.042h44min Mean–22.5126.2924.3726.432h51min –✓22.6625.7523.2424.392h44min Mean✓22.67 28.09 24.41 26.642h51min Max/Attn ✓22.6027.9724.3026.532h52min full model achieves the best PSNR on all scenes, demonstrating complementary gains. Replacing mean pooling in CLCA with max or attention yields comparable or slightly lower PSNR, validating mean aggregation as a sufficient regional prior. Training overhead is negligible: the full model requires only 2 additional minutes over the baseline. 5 Conclusion We presented HiCo-GS, a high-fidelity reconstruction framework that addresses two complementary limitations in octree-based ur- ban scene reconstruction. CLCA resolves cross-level feature isola- tion by leveraging the octree’s spatial containment structure for bidirectional hierarchical context flow, enriching fine-level fea- tures with structural priors and coarse-level features with detail statistics. DNGC enforces agreement between rendered and depth- derived normals through edge-aware smoothness losses, exploiting urban planar priors to suppress floating artifacts. We also intro- duced China-Pagoda, a benchmark of 8 ancient pagodas with over 30,000 images featuring dense ornamental details and repetitive textures. Experiments on Mill19, UrbanScene3D, MatrixCity, and China-Pagoda demonstrate state-of-the-art rendering quality and cleaner geometry across real-world and synthetic benchmarks. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Acknowledgments This work was supported by the National Natural Science Founda- tion of China under Grants 62571437 and 62471394. References [1] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. 2024. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics 31, 9 (2024), 6100–6111. [2] Yu Chen and Gim Hee Lee. 2024. Dogs: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus. Advances in Neural Information Processing Systems 37 (2024), 34487–34512. [3]Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. 2024. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 811–820. [4]Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. 2024. Swag: Splatting in the wild images with appearance-conditioned gaussians. In European Conference on Computer Vision. Springer, 325–340. [5] Abdelrhman Elrawy and Emad A Mohammed. 2025. Opacity-Gradient Driven Density Control for Compact and Efficient Few-Shot 3D Gaussian Splatting. arXiv preprint arXiv:2510.10257 (2025). [6]Jixuan Fan, Wanhua Li, Yifei Han, Tianru Dai, and Yansong Tang. 2025. Momentum-GS: Momentum Gaussian self-distillation for high-quality large scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Com- puter Vision. 25250–25260. [7]Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. 2024. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural information processing systems 37 (2024), 140138–140158. [8]Yuanyuan Gao, Hao Li, Jiaqi Chen, Zhengyu Zou, Zhihang Zhong, Dingwen Zhang, Xiao Sun, and Junwei Han. 2025. Citygs-x: A scalable architecture for effi- cient and geometrically accurate large-scale scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 27187–27196. [9] Antoine Guédon and Vincent Lepetit. 2024. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5354–5363. [10] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2024. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers. 1–11. [11] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. 2024. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5322–5332. [12]Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al. 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42, 4 (2023), 139–1. [13] Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. 2024. A hierarchical 3d gaussian repre- sentation for real-time rendering of very large datasets. ACM Transactions On Graphics (TOG) 43, 4 (2024), 1–15. [14]Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc Pollefeys, and Torsten Sattler. 2024. Wildgaussians: 3d gaussian splatting in the wild. arXiv preprint arXiv:2407.08447 (2024). [15]Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. 2023. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 3205–3215. [16]Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. 2023. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8456–8465. [17]Zhuoxiao Li, Shanliang Yao, Taoyu Wu, Yong Yue, Wufan Zhao, Rongjun Qin, An- gel F Garcia-Fernandez, Andrew Levers, and Xiaohui Zhu. 2024. ULSR-GS: ultra large-scale surface reconstruction Gaussian splatting with multi-view geometric consistency. arXiv preprint arXiv:2412.01402 (2024). [18] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al.2024. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5166–5175. [19]Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. 2022. Capturing, reconstructing, and simulating: the urbanscene3d dataset. In European Conference on Computer Vision. Springer, 93–109. [20]Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. 2024. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision. Springer, 265–282. [21]Yang Liu, Chuanchen Luo, Zhongkai Mao, Junran Peng, and Zhaoxiang Zhang. 2024. Citygaussianv2: Efficient and geometrically accurate reconstruction for large-scale scenes. arXiv preprint arXiv:2411.00771 (2024). [22]Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20654–20664. [23] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. 2024. Com- pressed 3d gaussian splatting for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10349– 10358. [24] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017). [25]Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. 2024. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898 (2024). [26]Yunji Seo, Young Sun Choi, Hyun Seung Son, and Youngjung Uh. 2024. Flod: Integrating flexible level of detail into 3d gaussian splatting for customizable rendering. arXiv preprint arXiv:2408.12894 (2024). [27] Jianxiong Shen, Yue Qian, and Xiaohang Zhan. 2025. Lod-gs: Achieving levels of detail using scalable gaussian soup. In Proceedings of the Computer Vision and Pattern Recognition Conference. 671–680. [28]Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. 2022. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12922–12931. [29]Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. 2025. Dn-splatter: Depth and normal priors for gaussian splatting and meshing. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2421–2431. [30] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. 2021. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2021). [31] Yongchang Wu, Zipeng Qi, Zhenwei Shi, and Zhengxia Zou. 2025. Blockgaussian: Efficient large-scale scene novel view synthesis via adaptive block-based gaussian splatting. arXiv preprint arXiv:2504.09048 (2025). [32]Haihong Xiao, Jianan Zou, Yuxin Zhou, and Wenxiong Kang. 2025. SplatCo: Structure-View Collaborative Gaussian Splatting for Detail-Preserving Rendering of Large-Scale Unbounded Scenes. arXiv preprint arXiv:2505.17951 (2025). [33] Jiawei Xu, Zexin Fan, Jian Yang, and Jin Xie. 2024. Grid4d: 4d decomposed hash encoding for high-fidelity dynamic gaussian splatting. Advances in Neural Information Processing Systems 37 (2024), 123787–123811. [34] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. 2024. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20923–20931. [35]Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 19447–19456. [36]Zehao Yu, Torsten Sattler, and Andreas Geiger. 2024. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (ToG) 43, 6 (2024), 1–13. [37]Wei Zhang, Qiang Li, and Qi Wang. 2025. Refined Cascade Cost Volume for Multi- View Remote Sensing Image Reconstruction. IEEE Transactions on Geoscience and Remote Sensing (2025). [38] Wei Zhang, Qiang Li, Yuan Yuan, and Qi Wang. 2024. Visual consistency enhance- ment for multiview stereo reconstruction in remote sensing. IEEE Transactions on Geoscience and Remote Sensing 62 (2024), 1–11. [39]Wei Zhang, Yihang Wu, Songhua Li, Wenjie Ma, Xin Ma, Qiang Li, and Qi Wang. 2025. Review of Feed-forward 3D Reconstruction: From DUSt3R to VGGT. arXiv:2507.08448 [cs.CV] https://arxiv.org/abs/2507.08448 [40]Wei Zhang, Yihang Wu, Shengkai Yu, Songhua Li, Qiang Li, and Qi Wang. 2026. GPR-MVS: Global Propagation Regularization for Large Scale Multi-view Stereo. IEEE Transactions on Geoscience and Remote Sensing (2026), 1–1. doi:10.1109/ TGRS.2026.3710063 [41]Wei Zhang, Zhigang Yang, Qiang Li, and Qi Wang. 2025. Semantic-guided Multi- view Stereo Reconstruction for Aerial Image. IEEE Transactions on Geoscience and Remote Sensing (2025). [42] Bingxuan Zhao, Chuang Yang, Qing Zhou, and Qi Wang. 2025. RLI-DM: Robust Layout-Based Iterative Diffusion Model for SAR-to-RGB Image Translation. IEEE Transactions on Geoscience and Remote Sensing (2025). [43]Bingxuan Zhao, Qing Zhou, Yu Wang, Chuang Yang, and Qi Wang. [n. d.].\Σ: Sig- moid Modulation for Ultra High Resolution Diffusion. In Forty-third International Conference on Machine Learning. [44]Brent Zoomers, Maarten Wijnants, Ivan Molenaers, Joni Vanherck, Jeroen Put, Lode Jorissen, and Nick Michiels. 2025. Progs: Progressive rendering of gaussian splats. In Proceedings of the Winter Conference on Applications of Computer Vision. 3118–3127. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Supplementary File I Cross-Level Feature Incoherence Analysis We provide a comprehensive feature-space analysis to empirically validate the cross-level feature isolation problem identified in the main paper and to characterize how CLCA reshapes the learned representations. The analysis spans four urban scenes: Building (8 octree levels), Rubble (5 levels), Residence (6 levels), and SciArt (7 levels). All measurements are taken from fully converged models at 100K iterations. A Cross-Level Cosine Distance Matrix For each model we extract all anchor featuresf 푖 and their oc- tree levels푙 푖 , then compute the mean pairwise cosine distance between every level pair(푙 푎 ,푙 푏 ): 푑(푙 푎 ,푙 푏 )= 1− 1 |A 푎 ||A 푏 | ∑︁ 푖∈A 푎 ∑︁ 푗∈A 푏 f ⊤ 푖 f 푗 ∥f 푖 ∥f 푗 ∥ ,(1) whereA 푙 =푖:푙 푖 =푙. A value of 1.0 indicates perfect orthogonal- ity. Baseline. Across all four scenes (Figs. 1–4, panel (a)), every off- diagonal entry falls within 0.94–1.00, confirming that baseline an- chor features at different octree levels are nearly orthogonal regard- less of their spatial relationship. The isolation worsens monotoni- cally with level depth: in Building, adjacent-level distances increase from 0.949 (L0↔L1) to 0.999 (L6↔L7), indicating that the finest levels suffer the most severe isolation. After CLCA. Panels (b) of the same images show that CLCA in- duces a block-diagonal structure. Fine-level features become highly coherent with one another (e.g. L5↔L6 drops from 0.995 to 0.343 in Building, a 65.5% reduction), while coarse-level features retain their original structure. This demonstrates a transition from unstructured orthogonality to hierarchically organized specialization. B Adjacent-Level Distance Reduction Panels (c) of Figs. 1–4 compare adjacent-level cosine distances be- fore and after CLCA. A consistent pattern emerges across all scenes: the reduction concentrates overwhelmingly on the finest level pairs. •Building (8 levels): L0↔L1 through L4↔L5 change by≤3.9%, while L5↔L6 and L6↔L7 are reduced by 65.5% and 25.0%. • Rubble (5 levels): L3↔L4 is reduced by 28.8%. •Residence (6 levels): L3↔L4 and L4↔L5 are reduced by 10.3% and 40.1%. •SciArt (7 levels): L4↔L5 and L5↔L6 are reduced by 46.1% and 62.9%. This adaptive behavior is not manually designed but emerges from training dynamics. Coarse levels contain many anchors whose mean-pooled context approximates the global average and carries little discriminative information. Fine levels contain few anchors with distinctive statistics; the parent-level context provides substan- tial structural information that independent optimization cannot access, producing large feature adjustments. The network thus au- tomatically concentrates coupling effort where incoherence is most severe. C Per-Pair PCA Analysis Global dimensionality reduction is dominated by coarse-level an- chors, which vastly outnumber fine-level ones. To reveal fine-grained cross-level structure, we perform per-pair PCA: for each adjacent pair(푙,푙+1), we extract only anchors at these two levels (subsam- pled to 3,000 per level), fit PCA on their combined features, and project to 2D. The cosine similarity between the two levels’ mean feature vectors is annotated. Fig. 1(d) shows the Building scene results. In the baseline (top row), most pairs exhibit overlapping point clouds (cos_sim≥0.78), confirming diffuse distributions without level-specific organization. After CLCA (bottom row), three distinct regimes appear: (i) coarse pairs remain unchanged; (i) at the coarse-to-fine transition (e.g. L4 vs L5), cos_sim drops sharply from 0.979 to 0.066 and the two levels form clearly separated clusters, with PC1 explained variance rising from 16.8% to 33.1%, indicating that the principal axis of variation has become a scale-discriminative direction; (i) at the finest pairs (e.g. L6 vs L7), both levels form compact, well-separated groups. This progression directly visualizes how CLCA induces functional specialization at fine octree levels. D t-SNE Feature Visualization To provide a holistic view of the entire feature space, we perform t-SNE on all levels simultaneously for each scene. Since fine levels contain far fewer anchors than coarse levels, we adopt stratified sampling to ensure adequate representation of every level. We use a perplexity of 30 and run for 1,000 iterations. Fig. 5 presents the results across all four scenes. In the baseline (left panels), every scene exhibits a single un- structured blob in which all levels are uniformly intermixed. No level-specific clustering is visible, confirming the absence of orga- nized inter-level structure: the near-orthogonality revealed by the cosine distance matrices manifests as a diffuse, undifferentiated embedding. After CLCA (right panels), the feature space self-organizes into a hierarchically structured layout. Coarse-level anchors (shown in cooler tones) remain intermixed with each other, forming a shared core that encodes global scene structure. Fine-level anchors, in contrast, separate into distinct, tight clusters at the periphery of the embedding. This pattern is strikingly consistent across scenes of varying depth and complexity: •Building (8 levels): L5, L6, and L7 each form clearly isolated satellite clusters around the coarse-level core (L0–L4). •Rubble (5 levels): L3 and L4 pull away from the main body, with L4 forming the most compact cluster. • Residence (6 levels): L5 separates as a small but distinct group, while L4 begins to detach from the coarse-level mass. •SciArt (7 levels): L5 and L6 form peripheral clusters, with the remaining levels sharing a common central region. This hierarchical organization emerges purely from the render- ing loss. CLCA provides fine-level anchors with sufficient shared context to form coherent groups, while their separation from coarse levels gives the shared MLP decoder a clear signal for scale-dependent attribute prediction. The transition from an unstructured blob to a HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil L0L1L2L3L4L5L6L7 L0 L1 L2 L3 L4 L5 L6 L7 0.9510.9490.9510.9530.9540.9630.9971.001 0.9490.9450.9460.9480.9490.9600.9971.004 0.9510.9460.9450.9460.9470.9580.9961.004 0.9530.9480.9460.9470.9480.9580.9951.003 0.9540.9490.9470.9480.9480.9580.9951.002 0.9630.9600.9580.9580.9580.9640.9951.000 0.9970.9970.9960.9950.9950.9950.9960.999 1.0011.0041.0041.0031.0021.0000.9990.991 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 10.07.55.02.50.02.55.07.5 PC1 (11.6%) 6 4 2 0 2 4 6 8 PC2 (7.8%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L7 L0L1L2L3L4L5L6L7 L0 L1 L2 L3 L4 L5 L6 L7 0.9510.9470.9470.9490.9540.9640.9640.985 0.9470.9360.9370.9410.9420.9950.9930.988 0.9470.9370.9300.9270.9310.9700.9610.960 0.9490.9410.9270.9080.9100.9220.9060.911 0.9540.9420.9310.9100.8990.9720.9510.916 0.9640.9950.9700.9220.9720.2800.3430.804 0.9640.9930.9610.9060.9510.3430.3420.749 0.9850.9880.9600.9110.9160.8040.7490.766 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 15105051015 PC1 (11.6%) 5 0 5 10 PC2 (7.8%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L7 Cross-Level Feature Incoherence Analysis Building (a) Baseline L0L1L2L3L4L5L6L7 L0 L1 L2 L3 L4 L5 L6 L7 0.9510.9490.9510.9530.9540.9630.9971.001 0.9490.9450.9460.9480.9490.9600.9971.004 0.9510.9460.9450.9460.9470.9580.9961.004 0.9530.9480.9460.9470.9480.9580.9951.003 0.9540.9490.9470.9480.9480.9580.9951.002 0.9630.9600.9580.9580.9580.9640.9951.000 0.9970.9970.9960.9950.9950.9950.9960.999 1.0011.0041.0041.0031.0021.0000.9990.991 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 10.07.55.02.50.02.55.07.5 PC1 (11.6%) 6 4 2 0 2 4 6 8 PC2 (7.8%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L7 L0L1L2L3L4L5L6L7 L0 L1 L2 L3 L4 L5 L6 L7 0.9510.9470.9470.9490.9540.9640.9640.985 0.9470.9360.9370.9410.9420.9950.9930.988 0.9470.9370.9300.9270.9310.9700.9610.960 0.9490.9410.9270.9080.9100.9220.9060.911 0.9540.9420.9310.9100.8990.9720.9510.916 0.9640.9950.9700.9220.9720.2800.3430.804 0.9640.9930.9610.9060.9510.3430.3420.749 0.9850.9880.9600.9110.9160.8040.7490.766 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 15105051015 PC1 (11.6%) 5 0 5 10 PC2 (7.8%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L7 Cross-Level Feature Incoherence Analysis Building (b) Ours (+CLCA) L0L1L1L2L2L3L3L4L4L5L5L6L6L7 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Distance +0.2% +0.9% +2.0% +3.9% -1.4% +65.5% +25.0% Adjacent-Level Feature Distance Building Baseline Ours (c) Adjacent-level distance comparison 50510 PC1 (17.1%) 6 4 2 0 2 4 6 PC2 (6.8%) L0 vs L1 Baseline (cos_sim=0.980) L0 L1 7.55.02.50.02.55.07.5 PC1 (17.3%) 6 4 2 0 2 4 6 PC2 (6.7%) L1 vs L2 Baseline (cos_sim=0.981) L1 L2 50510 PC1 (17.8%) 6 4 2 0 2 4 6 PC2 (6.6%) L2 vs L3 Baseline (cos_sim=0.995) L2 L3 1050510 PC1 (17.5%) 6 4 2 0 2 4 6 PC2 (6.5%) L3 vs L4 Baseline (cos_sim=0.996) L3 L4 50510 PC1 (16.8%) 6 4 2 0 2 4 6 8 PC2 (6.7%) L4 vs L5 Baseline (cos_sim=0.979) L4 L5 50510 PC1 (21.6%) 4 2 0 2 4 6 8 PC2 (9.6%) L6 vs L7 Baseline (cos_sim=0.782) L6 L7 1001020 PC1 (25.9%) 7.5 5.0 2.5 0.0 2.5 5.0 7.5 10.0 PC2 (9.6%) Ours (cos_sim=0.928) L0 L1 15105051015 PC1 (17.7%) 10 5 0 5 10 PC2 (7.6%) Ours (cos_sim=0.935) L1 L2 1050510 PC1 (15.4%) 10 5 0 5 10 15 20 PC2 (6.7%) Ours (cos_sim=0.910) L2 L3 105051015 PC1 (14.9%) 10 5 0 5 10 15 20 PC2 (7.3%) Ours (cos_sim=0.929) L3 L4 7.55.02.50.02.55.07.510.0 PC1 (33.1%) 15 10 5 0 5 10 15 20 PC2 (10.2%) Ours (cos_sim=0.066) L4 L5 5.02.50.02.55.07.5 PC1 (22.0%) 7.5 5.0 2.5 0.0 2.5 5.0 7.5 10.0 PC2 (9.5%) Ours (cos_sim=0.626) L6 L7 Per-Pair PCA Analysis Building (d) Per-pair PCA projections of adjacent-level features Figure 1: Cross-level feature analysis on the Building scene (8 octree levels, 5.98M anchors). (a) Baseline cosine distance matrix: all off-diagonal entries fall within 0.94–1.00, with incoherence worsening at finer levels (L6↔L7: 0.999). (b) After CLCA: a block- diagonal structure emerges, with fine-level coherence dramatically improved (L5↔L6: 0.995→0.343). (c) Adjacent-level distance reductions concentrate on the finest pairs (L5↔L6:−65.5%, L6↔L7:−25.0%), while coarse pairs change by≤4%. (d) Per-pair PCA: baseline features (top) are diffusely mixed across levels; after CLCA (bottom), fine-level pairs form clearly separated clusters with increased PC1 explained variance, indicating emergent scale-discriminative structure. hierarchically organized embedding is the most direct visual evi- dence that CLCA transforms the feature space from uninformative orthogonality into functional specialization. I Additional DNGC Ablation Results We present qualitative ablations of the DNGC regularization on the Building and Residence scenes. Figs. 6 and 7 compare normal maps and depth maps with and without DNGC. I China-Pagoda Dataset Details The China-Pagoda dataset contains 8 ancient Chinese pagodas, each captured with over 1,200 drone and ground-level images. Figs. 8 and 9 present sample images illustrating the diversity of archi- tectural styles and the challenging visual characteristics: dense ornamental carvings, curved multi-layer eaves, repetitive brick textures, and weathered stone surfaces. Figs. 10–17 show detailed reconstruction results including textured renders, extracted meshes, and top-down views. IV Large-Scale Urban Scene Comparison Figs. 18 and 19 compare HiCo-GS against CityGS-Xon the Urban- Scene3D Residence and Building scenes with full-scene textured renderings and extracted meshes. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. L0L1L2L3L4 L0 L1 L2 L3 L4 0.9700.9750.9770.9881.004 0.9750.9780.9780.9871.000 0.9770.9780.9780.9850.998 0.9880.9870.9850.9860.995 1.0041.0000.9980.9950.989 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 6420246 PC1 (11.2%) 8 6 4 2 0 2 4 6 PC2 (8.2%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L0L1L2L3L4 L0 L1 L2 L3 L4 0.9450.9450.9421.0041.094 0.9450.9430.9401.0041.093 0.9420.9400.9351.0081.110 1.0041.0041.0080.8470.708 1.0941.0931.1100.7080.251 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 12.510.07.55.02.50.02.55.07.5 PC1 (11.2%) 10.0 7.5 5.0 2.5 0.0 2.5 5.0 7.5 10.0 PC2 (8.2%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 Cross-Level Feature Incoherence Analysis Rubble (a) Baseline L0L1L2L3L4 L0 L1 L2 L3 L4 0.9700.9750.9770.9881.004 0.9750.9780.9780.9871.000 0.9770.9780.9780.9850.998 0.9880.9870.9850.9860.995 1.0041.0000.9980.9950.989 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 6420246 PC1 (11.2%) 8 6 4 2 0 2 4 6 PC2 (8.2%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L0L1L2L3L4 L0 L1 L2 L3 L4 0.9450.9450.9421.0041.094 0.9450.9430.9401.0041.093 0.9420.9400.9351.0081.110 1.0041.0041.0080.8470.708 1.0941.0931.1100.7080.251 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 12.510.07.55.02.50.02.55.07.5 PC1 (11.2%) 10.0 7.5 5.0 2.5 0.0 2.5 5.0 7.5 10.0 PC2 (8.2%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 Cross-Level Feature Incoherence Analysis Rubble (b) Ours (+CLCA) L0L1L1L2L2L3L3L4 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Distance +3.1% +3.9% -2.3% +28.8% Adjacent-Level Feature Distance Rubble Baseline Ours (c) Adjacent-level distance comparison 7.55.02.50.02.55.07.5 PC1 (14.5%) 6 4 2 0 2 4 6 PC2 (7.1%) L0 vs L1 Baseline (cos_sim=0.936) L0 L1 6420246 PC1 (15.7%) 4 2 0 2 4 6 PC2 (7.3%) L1 vs L2 Baseline (cos_sim=0.989) L1 L2 7.55.02.50.02.55.07.5 PC1 (15.4%) 6 4 2 0 2 4 6 PC2 (7.5%) L2 vs L3 Baseline (cos_sim=0.873) L2 L3 7.55.02.50.02.55.07.5 PC1 (17.8%) 4 2 0 2 4 6 PC2 (7.1%) L3 vs L4 Baseline (cos_sim=0.726) L3 L4 1050510 PC1 (18.2%) 8 6 4 2 0 2 4 6 8 PC2 (6.1%) Ours (cos_sim=0.983) L0 L1 1050510 PC1 (16.5%) 6 4 2 0 2 4 6 8 PC2 (5.8%) Ours (cos_sim=0.984) L1 L2 1510505101520 PC1 (15.5%) 10 5 0 5 10 PC2 (9.1%) Ours (cos_sim=-0.057) L2 L3 105051015 PC1 (31.2%) 10 5 0 5 10 15 PC2 (11.3%) Ours (cos_sim=0.838) L3 L4 Per-Pair PCA Analysis Rubble (d) Per-pair PCA projections of adjacent-level features Figure 2: Cross-level feature analysis on the Rubble scene (5 octree levels). The same analysis protocol as Fig. 1. The baseline exhibits uniform near-orthogonality across all level pairs. After CLCA, the finest pair (L3↔L4) shows a 28.8% distance reduction, while coarse pairs remain stable. Per-pair PCA confirms that CLCA selectively organizes fine-level features into separable clusters without disrupting coarse-level representations. L0L1L2L3L4L5 L0 L1 L2 L3 L4 L5 0.9670.9680.9700.9801.0111.066 0.9680.9680.9700.9801.0111.066 0.9700.9700.9720.9811.0091.059 0.9800.9800.9810.9851.0061.039 1.0111.0111.0091.0060.9910.969 1.0661.0661.0591.0390.9690.841 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 864202468 PC1 (15.6%) 8 6 4 2 0 2 4 6 8 PC2 (8.9%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L0L1L2L3L4L5 L0 L1 L2 L3 L4 L5 0.9730.9740.9740.9760.9830.998 0.9740.9720.9700.9730.9841.001 0.9740.9700.9630.9700.9891.011 0.9760.9730.9700.9330.9020.859 0.9830.9840.9890.9020.7270.580 0.9981.0011.0110.8590.5800.258 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 15105051015 PC1 (15.6%) 5 0 5 10 PC2 (8.9%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 Cross-Level Feature Incoherence Analysis Residence (a) Baseline L0L1L2L3L4L5 L0 L1 L2 L3 L4 L5 0.9670.9680.9700.9801.0111.066 0.9680.9680.9700.9801.0111.066 0.9700.9700.9720.9811.0091.059 0.9800.9800.9810.9851.0061.039 1.0111.0111.0091.0060.9910.969 1.0661.0661.0591.0390.9690.841 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 864202468 PC1 (15.6%) 8 6 4 2 0 2 4 6 8 PC2 (8.9%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L0L1L2L3L4L5 L0 L1 L2 L3 L4 L5 0.9730.9740.9740.9760.9830.998 0.9740.9720.9700.9730.9841.001 0.9740.9700.9630.9700.9891.011 0.9760.9730.9700.9330.9020.859 0.9830.9840.9890.9020.7270.580 0.9981.0011.0110.8590.5800.258 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 15105051015 PC1 (15.6%) 5 0 5 10 PC2 (8.9%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 Cross-Level Feature Incoherence Analysis Residence (b) Ours (+CLCA) L0L1L1L2L2L3L3L4L4L5 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Distance -0.6% +0.0% +1.1% +10.3% +40.1% Adjacent-Level Feature Distance Residence Baseline Ours (c) Adjacent-level distance comparison Figure 3: Cross-level feature analysis on the Residence scene (6 octree levels). CLCA reduces adjacent-level distances by up to 40.1% at the finest pair (L4↔L5), with coarse pairs showing minimal change. The adaptive coupling pattern is consistent with the other scenes. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil L0L1L2L3L4L5L6 L0 L1 L2 L3 L4 L5 L6 0.9370.9440.9490.9520.9680.9810.999 0.9440.9480.9530.9540.9700.9800.997 0.9490.9530.9560.9580.9720.9820.998 0.9520.9540.9580.9580.9720.9800.997 0.9680.9700.9720.9720.9800.9870.998 0.9810.9800.9820.9800.9870.9860.997 0.9990.9970.9980.9970.9980.9970.996 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 12.510.07.55.02.50.02.55.0 PC1 (20.2%) 8 6 4 2 0 2 4 6 PC2 (9.7%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L0L1L2L3L4L5L6 L0 L1 L2 L3 L4 L5 L6 0.9680.9700.9730.9710.9680.9540.984 0.9700.9710.9740.9720.9730.9600.990 0.9730.9740.9720.9710.9620.9480.972 0.9710.9720.9710.9630.9530.9350.971 0.9680.9730.9620.9530.6360.5320.517 0.9540.9600.9480.9350.5320.3720.370 0.9840.9900.9720.9710.5170.3700.281 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 3020100102030 PC1 (20.2%) 10 5 0 5 10 15 20 PC2 (9.7%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 Cross-Level Feature Incoherence Analysis SciArt (a) Baseline L0L1L2L3L4L5L6 L0 L1 L2 L3 L4 L5 L6 0.9370.9440.9490.9520.9680.9810.999 0.9440.9480.9530.9540.9700.9800.997 0.9490.9530.9560.9580.9720.9820.998 0.9520.9540.9580.9580.9720.9800.997 0.9680.9700.9720.9720.9800.9870.998 0.9810.9800.9820.9800.9870.9860.997 0.9990.9970.9980.9970.9980.9970.996 (a) Baseline: Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 12.510.07.55.02.50.02.55.0 PC1 (20.2%) 8 6 4 2 0 2 4 6 PC2 (9.7%) (b) Baseline: PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 L0L1L2L3L4L5L6 L0 L1 L2 L3 L4 L5 L6 0.9680.9700.9730.9710.9680.9540.984 0.9700.9710.9740.9720.9730.9600.990 0.9730.9740.9720.9710.9620.9480.972 0.9710.9720.9710.9630.9530.9350.971 0.9680.9730.9620.9530.6360.5320.517 0.9540.9600.9480.9350.5320.3720.370 0.9840.9900.9720.9710.5170.3700.281 (c) Ours (after CLCA): Cross-Level Cosine Distance 0.0 0.2 0.4 0.6 0.8 1.0 3020100102030 PC1 (20.2%) 10 5 0 5 10 15 20 PC2 (9.7%) (d) Ours (after CLCA): PCA Feature Distribution L0 L1 L2 L3 L4 L5 L6 Cross-Level Feature Incoherence Analysis SciArt (b) Ours (+CLCA) L0L1L1L2L2L3L3L4L4L5L5L6 0.0 0.2 0.4 0.6 0.8 1.0 Cosine Distance -2.8% -2.2% -1.4% +2.0% +46.1% +62.9% Adjacent-Level Feature Distance SciArt Baseline Ours (c) Adjacent-level distance comparison Figure 4: Cross-level feature analysis on the SciArt scene (7 octree levels). SciArt exhibits the largest fine-level reductions among all scenes (L5↔L6:−62.9%), consistent with its higher geometric complexity requiring stronger cross-level communication. Baseline L0 (1500) L1 (1500) L2 (1500) L3 (1500) L4 (1500) L5 (1500) L6 (1500) L7 (1500) Ours (after CLCA) L0 (1500) L1 (1500) L2 (1500) L3 (1500) L4 (1500) L5 (1500) L6 (1500) L7 (1500) t-SNE Feature Visualization Building (a) Building (8 levels, 1,500 per level) Baseline L0 (2400) L1 (2400) L2 (2400) L3 (2400) L4 (2400) Ours (after CLCA) L0 (2400) L1 (2400) L2 (2400) L3 (2400) L4 (2400) t-SNE Feature Visualization Rubble (b) Rubble (5 levels, 2,400 per level) Baseline L0 (2000) L1 (2000) L2 (2000) L3 (2000) L4 (2000) L5 (122) Ours (after CLCA) L0 (2000) L1 (2000) L2 (2000) L3 (2000) L4 (2000) L5 (55) t-SNE Feature Visualization Residence (c) Residence (6 levels, 2,000 per level) Baseline L0 (1714) L1 (1714) L2 (1714) L3 (1714) L4 (1714) L5 (1714) L6 (1714) Ours (after CLCA) L0 (1714) L1 (1714) L2 (1714) L3 (1714) L4 (1714) L5 (1714) L6 (1714) t-SNE Feature Visualization SciArt (d) SciArt (7 levels, 1,714 per level) Figure 5: t-SNE visualization of anchor features across all octree levels. Each panel shows the baseline (left) and after CLCA (right) for one scene, with stratified sampling ensuring equal representation per level. In the baseline, all levels form a single unstructured mixture. After CLCA, fine-level features consistently separate into distinct peripheral clusters while coarse levels share a common core, demonstrating the emergent hierarchical specialization induced by cross-level context aggregation. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Figure 6: DNGC ablation on the Building scene. Top: With DNGC, rendered normals exhibit clean, consistent orientations on planar surfaces such as rooftops and walls, and depth maps show smooth gradients with sharp object boundaries. Bottom: Without DNGC, normals are noticeably noisier with inconsistent directions on the same regions, and depth maps contain discontinuities and floating artifacts around building facades and ground planes. Figure 7: DNGC ablation on the Residence scene. Top: Without DNGC, normal maps show fragmented orientations on rooftops and walls, with depth artifacts at structural boundaries. Bottom: With DNGC, normals become coherent on large planar surfaces, and depth transitions are cleaner at building edges. The edge-aware smoothness term preserves sharp boundaries between adjacent structures while enforcing planarity within each surface. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Figure 8: China-Pagoda sample images (Part 1): Beita, Duobaota, Lingshan, and Liuhe pagodas shown from multiple drone viewpoints. These pagodas feature octagonal and hexagonal cross-sections, tiered eaves with hanging bells, intricate brick carvings, and arched doorways. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Figure 9: China-Pagoda sample images (Part 2): Yunjusi, Yongwangta, Yuhuangta, and Baoanta pagodas. These structures exhibit heavily weathered surfaces, dense repetitive textures from layered brick construction, and fine-grained ornamental details at varying scales. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Duobaota TextureMesh Figure 10: Reconstruction of Duobaota. From left to right: textured rendering, extracted mesh, and top-down views. The multi-layer eaves, hanging bells, and hexagonal base geometry are faithfully reconstructed with clean surface normals and sharp edge boundaries. Yunjusi TextureMesh Figure 11: Reconstruction of Yunjusi. The tallest pagoda in our dataset with seven tiers of eaves and dense window openings. The mesh preserves octagonal symmetry and layered eave overhangs, with top-down views revealing clean rooftop geometry and radial ridge lines. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Lingshan TextureMesh Figure 12: Reconstruction of Lingshan. This pagoda features prominent curved eaves with decorative bells and a surrounding stone balustrade. The mesh accurately captures eave curvature, the octagonal tiered structure, and fine railing geometry at the base. Baoanta Figure 13: Reconstruction of Baoanta. Textured rendering, extracted mesh, and top-down views demonstrating faithful recovery of the pagoda’s layered structure and surface detail. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Beita Figure 14: Reconstruction of Beita. The dense brick carvings and weathered surface patterns are well preserved in both the rendered appearance and the extracted mesh geometry. Liuheta Figure 15: Reconstruction of Liuheta. The multi-story structure with alternating eave tiers and window openings is reconstructed with consistent geometry across all levels. M ’26, November 10–14, 2026, Rio de Janeiro, BrazilZhang et al. Yongwangta Figure 16: Reconstruction of Yongwangta. Clean mesh extraction captures the tapered silhouette and tiered eave overhangs characteristic of this architectural style. Yuhuangta Figure 17: Reconstruction of Yuhuangta. Fine ornamental details and the glazed-tile surface texture are faithfully reproduced across multiple viewpoints. HiCo-GS: Hierarchical Context Aggregation and Geometric Consistency for Octree Gaussian SplattingMM ’26, November 10–14, 2026, Rio de Janeiro, Brazil HiCo-GS (Residence) CityGS-X (Residence) TextureMesh TextureMesh Figure 18: Full-scene comparison on Residence. Top: HiCo-GS produces consistent facade coloring and clean planar surfaces on rooftops and roads. Bottom: CityGS-Xexhibits more floating artifacts and noisier mesh geometry, especially around vegetation boundaries and overpass structures. HiCo-GS (Building) CityGS-X (Building) TextureMesh TextureMesh Figure 19: Full-scene comparison on Building. Top: HiCo-GS achieves well-defined building edges and smooth ground planes. Bottom: CityGS-Xshows more surface noise on large flat regions such as the sports field and rooftops, where the lack of cross-level context and geometric consistency leads to less coherent surface reconstruction.