Paper deep dive
Standalone DINOv3 for Training-Free Open-Vocabulary Semantic Segmentation in Remote Sensing
Changhao Zhao, Haoxiang Li, Yuke Li, Hai Liu, LingLin Zeng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/5/2026, 4:51:09 AM
Summary
The paper introduces DinoSplat-OV, a training-free open-vocabulary semantic segmentation framework for remote sensing imagery based on the DINOv3 backbone. It addresses the challenges of dense object distribution, multi-scale targets, and large image sizes without fine-tuning. Key components include Text-aware Laplacian Propagation (TLP) for denoising patch-level predictions, Gaussian Splatting Upsampling (GSUP) for pixel-level feature reconstruction, and a global-anchor sliding-window strategy for large-scale inference. The method demonstrates competitive performance on datasets like UDD5, DOTA, and LoveDA.
Entities (11)
Relation Signals (9)
DinoSplat-OV â usesbackbone â DINOv3
confidence 95% · We propose DinoSplat-OV, a training-free framework that adapts DINOv3 to remote sensing
DinoSplat-OV â usesmodule â Gaussian Splatting Upsampling
confidence 95% · Its Gaussian Splatting Upsampling module(GSUP) reconstructs pixel-level features
DinoSplat-OV â usesmodule â Text-aware Laplacian Propagation
confidence 95% · Its Text-aware Laplacian Propagation module(TLP) de-noises patch-level predictions
DINOv3 â pairedwith â DINO.txt
confidence 92% · DINO.txt(Jose et al. 2024) is the text encoder paired with DINOv3
DinoSplat-OV â evaluatedon â DOTA
confidence 90% · Experiments on UDD5, DOTA, and LoveDA demonstrate competitive or superior performance
DinoSplat-OV â evaluatedon â UDD5
confidence 90% · Experiments on UDD5, DOTA, and LoveDA demonstrate competitive or superior performance
DinoSplat-OV â evaluatedon â LoveDA
confidence 90% · Experiments on UDD5, DOTA, and LoveDA demonstrate competitive or superior performance
DinoSplat-OV â evaluatedon â Vaihingen
confidence 90% · Experiments on UDD5, DOTA, LoveDA and Vaihingen demonstrate competitive or even superior performance
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Remote sensing semantic segmentation is hindered by costly pixel-level annotations, motivating training-free open-vocabulary methods. Recently, the recent release of DINOv3 brings this http URL, which equips the standalone DINO backbone with image-text contrastive learning and thus opens up the possibility of open-vocabulary segmentation. We propose DinoSplat-OV, a training-free framework that adapts DINOv3 to remote sensing without fine-tuning or additional pretraining. Targeting the dense distribution, multi-scale nature, and large size of remote sensing imagery, we design two core modules. Its Text-aware Laplacian Propagation module de-noises patch-level predictions by combining textual semantic affinities with local visual similarity, improving regional consistency while preserving boundaries. Its Gaussian Splatting Upsampling module reconstructs pixel-level features through RGB-guided anisotropic aggregation and test-time optimization. A global-anchor sliding-window strategy further supports large-scale imagery. Experiments on UDD5, DOTA, and LoveDA demonstrate competitive or superior performance over existing training-free methods, effectively filling the gap of DINO-series models in training-free open-vocabulary segmentation and providing a viable new path for further advances in this direction.
Tags
Links
- Source: https://arxiv.org/abs/2608.03023v1
- Canonical: https://arxiv.org/abs/2608.03023v1
Trouble viewing inline? Open PDF directly â
Full Text
38,887 characters extracted from source content.
Expand or collapse full text
Standalone DINOv3 for Remote Sensing Training-Free Open-Vocabulary Semantic Segmentation Changhao Zhao1, Haoxiang Li1, Yuke Li2, Hai Liu1, LingLin Zeng1,â Abstract Remote sensing semantic segmentation is hindered by costly pixel-level annotations, motivating training-free open-vocabulary methods. Recently, The recent release of DINOv3 brings DINO.txt, which equips the standalone DINO backbone with imageâtext contrastive learning and thus opens up the possibility of openâvocabulary segmentation.We propose DinoSplat-OV, a training-free framework that adapts DINOv3 to remote sensing without fine-tuning or additional pretraining. Targeting the dense distribution, multi-scale nature, and large size of remote sensing imagery, we design two core modules. Its Text-aware Laplacian Propagation module(TLP) de-noises patch-level predictions by combining textual semantic affinities with local visual similarity, improving regional consistency while preserving boundaries. Its Gaussian Splatting Upsampling module(GSUP) reconstructs pixel-level features through RGB-guided anisotropic aggregation and test-time optimization. A global-anchor sliding-window strategy further supports large-scale imagery. Experiments on UDD5, DOTA, LoveDA and Vaihingen demonstrate competitive or superior performance over existing training-free methods, effectively filling the gap of DINO-series models in training-free open-vocabulary segmentation and providing a viable new path for further advances in this direction. Introduction Remote sensing imagery plays a critical role in precision agriculture, disaster response, and environmental monitoring. However, the acquisition of remote sensing data and pixelâlevel annotation are extremely costly, and the images are characterized by dense objects, varying scales, and huge sizes, making vision foundation models (VFMs) preâtrained on natural images difficult to apply directly. Extensive prior work has attempted fineâtuningâbased adaptation, yet still faces generalization bottlenecks. Openâvocabulary semantic segmentation, which can recognize categories defined by arbitrary text, has emerged as the most promising direction for remote sensing interpretation. Existing methods mostly adopt CLIP as the backbone and postâprocess its logits to eliminate global biases, among which trainingâfree schemes have achieved notable progress. Nevertheless, such optimization heavily relies on CLIPâs dualâencoder architecture and modifiable internal modules, and cannot be directly transferred to other VFMs. The release of DINOv3 and its accompanying text encoder (DINO.txt) opens up the possibility of using DINO alone for openâvocabulary segmentation. Although DINOv3âs visual features are superior to those of CLIP, its visual backbone is completely frozen under the LiT training paradigm, and we are unable to modify its internal attention or feedâforward layers as we do with CLIP. This leads to severe noise and blurred boundaries when directly applied to remote sensing imagery. To address this, we systematically transfer the training free experience from CLIP to DINOv3 and propose DinoSplatâOV. This model is designed for the largeâscale and highâdensity characteristics of remote sensing, with a purely inferenceâoriented optimization pipeline: synonym aggregation alleviates textâimage matching fragmentation, Textâaware Laplacian Propagation (TLP) guides feature alignment, 2D Gaussian Splatting Upsampling (GSUP) reconstructs lowâresolution features to pixelâlevel precision, and globalâanchor sliding window supports arbitrary largeâimage inference. The contributions of this paper are threefold: âą We systematically analyze the differences between DINO and CLIP, and propose the first trainingâfree openâvocabulary segmentation framework for remote sensing based on DINOv3. âą We develop two core algorithms tailored to DINO.txt: the Text-aware graph Laplacian Propagation (TLP) for coarse-grained feature optimization, and the 2D Gaussian Splatting-inspired upsampling (GSUP) for accurate pixel-level feature reconstruction. âą We validate competitive or even superior performance to the stateâofâtheâart on UDD5, DOTA, LoveDA, Vaihingen and other datasets, without requiring a preâtrained upsampler. Related Work Training-Free Open-Vocabulary Segmentation Paradigms. Figure 1: DinoSplatâOV employs a slidingâwindow strategy to process large remote sensing images; after feature extraction by DINOv3, the features are sequentially refined via Textâaware Laplacian Propagation (TLP), upsampled through 2D Gaussian Splatting Upsampling (GSUP), and fused with globalâanchor window attention, ultimately producing pixelâlevel segmentation predictions. Open-vocabulary segmentation has progressed with vision foundation models (VFMs). CLIP, with its dual-encoder joint training, has become the dominant backbone. Recent training-free efforts, such as SCLIP (Wang et al. 2023b) (modifying self-attention to Query-Query) and ClearCLIP (Lan et al. 2024a) (removing FFN and residuals), focus on mitigating CLIPâs inherent global bias for dense prediction. However, these architectural modifications are intrinsically tied to CLIPâs modifiable internal modules. In parallel, while DINO offers superior visual features, existing works (e.g., clip-dinosier (WysoczaĆska et al. 2024), proxyCLIP (Lan et al. 2024b), LPOSS (StojniÄ et al. 2025)) merely treat it as an auxiliary denoising tool for CLIP, rather than an independent segmenter. The recent release of DINOv3 with its text encoder (DINO.txt) enables DINO-only segmentation for the first time. Critically, unlike CLIP, DINOv3 employs a Locked-image Tuning (LiT) paradigm that fully freezes the visual backbone, making the successful CLIP-style internal modifications inapplicable. This frozen constraint constitutes the core challenge our inference-time optimizations must address. Remote Sensing Adaptations and Domain Gaps. Remote sensing imagery presents unique hurdles: dense object distribution, multi-scale targets, and gigapixel resolutions. While OVRS(Cao et al. 2025) employs a cost-aggregation training scheme adapted from Cat-Seg(Cho et al. 2024) to remote sensing data., and SegEarth-OV/SatOV (Li et al. 2025) incorporate pre-trained upsamplers (e.g. FeatUp (Fu et al. 2024)) to recover resolution, these methods still rely on dataset-specific pre-training. Consequently, they suffer from degraded generalization when facing unseen geographic regions. Moreover, GLACLIP(Lee et al. 2026) introduces a sliding-window strategy tailored for remote sensing, which differs from the standard one. These domain-specific bottlenecksâhigh-resolution recovery and seamless large-image inferenceâare exacerbated when using a frozen DINO backbone, as we cannot fine-tune the features to adapt to RS distributions. Our Method Preliminaries DINOv3 Text Encoder (DINO.txt) DINO.txt(Jose et al. 2024) is the text encoder paired with DINOv3, enabling openâvocabulary segmentation using DINO alone. Its training paradigm differs fundamentally from CLIPâs joint training: DINO.txt adopts a Lockedâimage Tuning (LiT) strategy, where the visual backbone is fully frozen and only the text encoder is optimized for crossâmodal alignment. This design preserves the strong discriminative features learned during selfâsupervised preâtraining, but consequently prohibits modifying internal attention or feedâforward layers at inference time. In contrast, CLIPâs dualâencoder architecture allows flexible adjustments to its visual branch, highlighting a core architectural distinction between the two models. 3D Gaussian Splatting 3D Gaussian Splatting (3DGS) (Kerbl et al. 2023) represents continuous visual fields using explicit Gaussian primitives, where each primitive is parameterized by position, covariance, opacity, and appearance attributes. Through differentiable projection and alpha blending, 3DGS reconstructs dense signals from sparse primitives without relying on complex neural decoders. Recent work such as Feat2GS (Chen et al. 2025) further demonstrates that visual foundation model features can be effectively represented by Gaussian primitives. Inspired by this idea, we extend Gaussian splatting from visual rendering to semantic feature reconstruction, treating DINO tokens as semantic Gaussian primitives for high-resolution feature upsampling through adaptive weighted aggregation. Overall Architecture The inference pipeline of DinoSplat-OV (Figure 1) proceeds through four sequential modules. We first apply synonym aggregation on the text side, fusing embeddings from multiple synonymous descriptions to compensate for DINO.txtâs relatively weak cross-modal alignment. For the frozen visual features, Text-aware Laplacian Propagation (TLP) performs anisotropic diffusion on the initial logits under the guidance of text priors: it aggressively smooths semantically homogeneous regions to suppress noise and fill holes, while preserving discontinuities at land-cover boundaries. The resulting coarse logits are then fed into Gaussian Splatting Upsampling (GSUP), which treats each low-resolution pixel as a 2D Gaussian primitive and reconstructs pixel-level feature maps via RGB-guided anisotropic weightingâall through test-time optimization without any pre-trained upsampler. Finally, to handle gigapixel remote sensing inputs, a global-anchor sliding window strategy externally simulates global attention by using the CLS token as contextual reference across windows, coupled with Hann window weighting to eliminate stitching artifacts. This purely inference-oriented design requires neither fine-tuning nor retraining. Synonym Aggregation for Text Embeddings To compensate for the weak cross-modal alignment of DINO.txt, we aggregate multiple synonymous descriptions per category. For each class c, let c,ii=1Nc\t_c,i\_i=1^N_c be the text embeddings of its NcN_c synonymous phrases. We compute a unified class embedding via weighted averaging: c=âi=1Ncwc,iâc,iâi=1Ncwc,i,T_c= _i=1^N_cw_c,it_c,i _i=1^N_cw_c,i, (1) where wc,iw_c,i are set uniformly in our main experiments, but can be adjusted based on phrase frequency or importance. This aggregation reduces the variance caused by single-word expression biases, providing more stable text priors for subsequent modules. Text-aware Laplacian Propagation (TLP) Figure 2: Cost Aggregation vs Laplacian Propagation Diagram To mitigate noise and cross-window inconsistency in segmentation predictions on low-resolution feature maps, existing post-processing approachesâsuch as Cost Aggregation (shou2024cat)ârequire training on annotated data. To address this limitation, we propose a training-free graph Laplacian propagation algorithm, termed Text-aware Laplacian Propagation (TLP). This method is essentially a variant of label propagation and can be viewed as a closed-form simplification of graph convolutional networks (GCNs) when applied to a single image. Its core idea is to leverage text-embedding priors to construct global semantic relationships among categories and to couple these with local visual features of the image. This coupling guides the classification logits via anisotropic diffusionâsmoothing sufficiently within semantically homogeneous regions to eliminate noise, while suppressing cross-region propagation at semantic boundaries to preserve land-cover contours. Given the aggregated class embeddings ââCĂDT ^CĂ D obtained from Eq. (1), we first construct a semantic correlation matrix ââCĂCS ^CĂ C via cosine similarity with temperature scaling: Siâj=expâĄ(iâ€âj/ÏS)âk=1CexpâĄ(iâ€âk/ÏS),S_ij= (T_i T_j/ _S ) _k=1^C (T_i T_k/ _S ), (2) where ÏS _S is a temperature parameter.In practice, we further enhance the diagonal entries and perform row-wise renormalization. In our implementation, we enforce symmetry via Sâ(S+Sâ€)/2Sâ(S+S )/2 to maintain a balanced propagation prior. This matrix serves as a global prior to modulate the subsequent local diffusion intensity. Let PââBĂCĂHĂWP ^BĂ CĂ HĂ W be the probability map after softmax normalization, and let pcâ(u)p_c(u) denote the probability that position u belongs to category c. We define the prediction confidence at position u as Îłâ(u)=maxcâĄpcâ(u)Îł(u)= _cp_c(u), and the semantic self-consistency as αâ(u)=âcpcâ(u)ââjScâjâpjâ(u),α(u)= _cp_c(u) _jS_cjp_j(u), (3) which measures how well the pixelâs probability distribution agrees with the text semantic matrix Sâif the predicted category exhibits consistent responses with semantically similar categories, αâ(u)α(u) will be high. The diffusion gating coefficient is then defined as: λ(u)=maxÎł(u),Îłmin2â (1+α(u)),λ(u)= \Îł(u), _ \^2· (1+α(u) ), (4) where Îłmin _ is a small constant (set to 0.050.05 in experiments) to prevent numerical instability. λâ(u)λ(u) controls the diffusion strength: regions with high confidence and strong semantic self-consistency are sufficiently smoothed to eliminate noise, while regions with low confidence or semantic ambiguity suppress diffusion to preserve fine details. For adjacent pixel positions u and v, we define the anisotropic edge weight Ïu,vâ[0,1] _u,vâ[0,1], which integrates both image- and semantic-guided terms: Ïu,v=Ïu,vimâ Ïu,vsem,Ï^u,v=Ïu,v+Ïv,u2, _u,v= _u,v^im· _u,v^sem, Ï_u,v= _u,v+ _v,u2, (5) where Ïu,vimg=expâĄ(âkIâ |Iâ(u)âIâ(v)|ÎŒI), _u,v^img= \! (-k_I· |I(u)-I(v)| _I ), with ÎŒI=meanâ(|âI|) _I=mean(|â I|) being the mean absolute grayscale gradient over the local map and kIk_I a constant (set to 5.05.0 in experiments), and the semantic term is modulated as Ïu,vsem=1+âcpcâ(u)ââjScâjâpjâ(v). _u,v^sem=1+ _cp_c(u) _jS_cj\,p_j(v). In practice, we apply a lightweight numerical projection Ï^u,vâminâĄ(1,Ï^u,v) Ï_u,vâ (1, Ï_u,v) to strictly confine the symmetric edge weights within [0,1][0,1] as a stabilizer. Let XââBĂCĂHĂWX ^BĂ CĂ HĂ W denote the classification logits to be optimized (initialized as the input logits). Define the graph Laplacian operator âL acting on a feature map Z as: ââ(Z)â(u)=âvââ(u)Ï^u,vâ(Zâ(u)âZâ(v))L(Z)(u)= _v (u) Ï_u,v(Z(u)-Z(v)) (6) where â(u)N(u) denotes the four-neighborhood. The final smoothing process is accomplished by solving the following sparse linear system: λâ(u)âXâ(u)+Ïâ ââXâ(u)=λâ(u)â Xinâ(u),âu,λ(u)X(u)+Ï·LX(u)=λ(u)· X_in(u), â u, (7) where XinX_in is the original input logits and Ï>0Ï>0 is a smoothing strength parameter. This formulation is equivalent to performing anisotropic smoothing on low-frequency regions while preserving high-frequency boundaries. The raw edge weight Ïu,v _u,v defined in Eq. (5) is generally asymmetric due to the semantic term. To enable a symmetric graph Laplacian and safely apply the Conjugate Gradient (CG) method, we symmetrize it as Ï^u,v=(Ïu,v+Ïv,u)/2 Ï_u,v=( _u,v+ _v,u)/2 (Eq. (5a)). With this symmetrization, the coefficient matrix in Eq. (7) becomes symmetric positive definite (since λâ(u)>0λ(u)>0 and the graph Laplacian is positive semidefinite), thus guaranteeing the convergence of CG. We therefore employ the CG method for efficient iterative solution, which converges robustly within a fixed number of steps without requiring backpropagation or training. In practice, to accelerate inference on high-resolution images while maintaining global receptive fields, we downsample the input logits and image to a coarse grid (determined by a target size, e.g., 72Ă7272Ă 72), solve the linear system in Eq. (7) on this downsampled space, and upsample the solution back to the original resolution with GSUP. This downsampling strategy is also consistent with our implementation, where the TLP module operates on the patch tokensâ spatial grid before upsampling. GSUP: Gaussian Splatting Inspired Upsampling Figure 3: 3D Gaussian Splatting vs 2D Gaussian Splatting Upsampling Diagram For dense segmentation scenarios in remote sensing imagery, recovering lowâresolution features to pixelâlevel accuracy is of critical importance. Existing methods such as FeatUp and AnyUp(Wimmer et al. 2026) rely on preâtrained upsampling models. However, since their backbones are predominantly trained on natural image datasets like ImageNet, remote sensing data constitutes only a small proportion of their training distribution, leading to notable crossâdomain generalization bottlenecks. Recently, NAF(Chambon et al. 2025) has improved JBU by replacing its fixed kernel with neighborhood attention, while 3DGS has demonstrated powerful explicit scene representation via anisotropic covariance without pretraining, achieving highâquality reconstruction through testâtime optimization (TTO). Inspired by both, we propose a feature reconstruction upsampling module based on 2D Gaussian Splatting, termed Gaussian Splatting Upsampling (GSUP), which similarly replaces JBUâs fixed kernel with a Gaussian splatting kernel. Model Backbone UDD5 DOTA LoveDA Vaihingen Average MaskCLIP ECCV baseline CLIP 28.1 12.4 22.6 30.0 23.3 DINO.txt CVPR DINOv3 32.4 17.0 27.8 35.7 28.2 ClearCLIP ECCV CLIP 38.2 18.5 31.6 39.4 31.9 LPOSS CVPR CLIP + DINO 38.8 20.2 32.4 32.0 30.8 SegEarth-OV CVPR ClearCLIP 45.3 22.3 36.9 40.9 36.3 DinoSplat-OVOurs DINO.txt 42.9 28.6 36.3 42.3 37.5 Table 1: Comparison of backbones and mIoU for different training free methods. Figure 4: Visualization of segmentation results for different methods on different datasets. Formal Definition. Let the low-resolution feature map be lrââBĂCĂHlĂWlF_lr ^BĂ CĂ H_lĂ W_l, and the target high-resolution feature map be hrââBĂCĂHhĂWhF_hr ^BĂ CĂ H_hĂ W_h, where Hh/Hl=Wh/Wl=rH_h/H_l=W_h/W_l=r is the upsampling factor. We define each pixel qâlq _l on the low-resolution feature map (where lG_l denotes the HlĂWlH_lĂ W_l grid) as a 2D Gaussian splatting primitive, with its coordinate qââ2x_q ^2 serving as the spatial center, the corresponding feature vector qââCf_q ^C as the propagable attribute, and the anisotropic covariance matrix qââ2Ă2 _q ^2Ă 2 characterizing its spatial morphology. The reconstructed feature at a high-resolution target position pâhp _h (where hG_h denotes the HhĂWhH_hĂ W_h grid) is obtained by weighted aggregation of all low-resolution Gaussian primitives: hrâ(p)=âqââ(p)wp,qâ lrâ(q),F_hr(p)= _q (p)w_p,q·F_lr(q), (8) where â(p)N(p) is the set of spatially proximate low-resolution pixels to p (implemented via k-nearest neighbor truncation), and wp,qâ[0,1]w_p,qâ[0,1] are the normalized splatting weights satisfying âqââ(p)wp,q=1 _q (p)w_p,q=1. Anisotropic Adaptive Weights. Unlike JBU and similar methods that adopt fixed isotropic kernels, we leverage the current RGB image as guidance to generalize the weights wp,qw_p,q to an adaptive anisotropic Gaussian splatting kernel, composed of a spatial term and a color term: wp,q=Softmaxqââ(p)âĄ(Ïsâ(p,q)+Ïcâ(p,q)),w_p,q=Softmax_q (p) ( _s(p,q)+ _c(p,q) ), (9) where the spatial term employs Mahalanobis distance to measure geometric deformation: Ïsâ(p,q)=â12âÎâp,qâ€âqâ1âÎâp,q,Îâp,q=pâq, _s(p,q)=- 12 _p,q _q^-1 _p,q, _p,q=x_p-x_q, (10) with q _q being a learnable Gaussian covariance matrix, parameterized by its scale parameters q=(sx,sy)s_q=(s_x,s_y) and rotation angle Ξq _q: q=â(Ξq)â diagâĄ(sx2,sy2)â â(Ξq)†_q=R( _q)·diag(s_x^2,s_y^2)·R( _q) . The color term measures semantic similarity based on RGB pixel differences: Ïcâ(p,q)=âââ(p)ââ(q)â222âÏr,q2, _c(p,q)=- \|I(p)-I(q)\|_2^22 _r,q^2, (11) where â(p)ââ3I(p) ^3 is the RGB color value, and Ïr,q _r,q is a learnable color bandwidth parameter. Test-Time Optimization (TTO). In contrast to the feed-forward sampling of traditional JBU, our method treats the Gaussian parameters q,Ξq,Ïr,q\s_q, _q, _r,q\ as optimizable variables, takes the low-resolution RGB image lrI_lr as input and the original high-resolution RGB image hrI_hr as supervisory signal, and performs test-time optimization by minimizing the L1 reconstruction loss: âTTO=ââ±GSUPâ(lr;)âhrâ1,L_TTO= \|F_GSUP(I_lr; )-I_hr \|_1, (12) where â±GSUPâ(â ;)F_GSUP(·; ) denotes the forward splatting projection that takes the low-resolution RGB values as the "values", and =q,Ξq,Ïr,qqâl =\s_q, _q, _r,q\_q _l. This optimization involves only the Gaussian parameters and does not update any visual backbone or feature encoder. It requires only a few dozen SGD iterations (e.g., 10 steps) and a sparse neighbor set (K=16), which incurs <3 GB GPU memory and 1200 ms latency per image, incurring minimal inference overhead. After optimization, with fixed, the low-resolution semantic features lrF_lr are fed as the "values" into the same forward pass, yielding the pixel-level high-resolution semantic feature map hrF_hr. The entire pipeline requires no pre-training and relies entirely on test-time adaptation, providing a lightweight and generalizable alternative for high-resolution recovery from frozen backbone features. Global-Anchor Window Attention Vision foundation models such as CLIP and DINO are trained on images of fixed resolution (e.g., 224Ă224224Ă 224 pixels) and are highly sensitive to input sizes. Remote sensing images, however, typically exceed 1000Ă10001000Ă 1000 pixels, making sliding window strategies essential for processing large-scale imagery. Nevertheless, conventional independent window inference introduces severe stitching artifacts, often resulting in discontinuities at window boundaries and disrupting the spatial consistency of segmentation predictions. To address this, we propose a global-anchor guided sliding window attention mechanism that simulates global self-attention externally, enabling each window to perceive global semantic context during decision-making. Specifically, we adopt the CLS token from ViT as the global anchor globalââDF_global ^D, which aggregates the global representation of the current window during the window encoding process. For the feature iââDF_i ^D of the i-th window, the global-aware weight is computed as: The global-aware weight for the i-th window is computed as: wi=expâĄ(simâ(Fi,Fglobal)/Ï)âjexpâĄ(simâ(Fj,Fglobal)/Ï)â GÏâ(i),w_i= (sim(F_i,F_global)/Ï ) _j (sim(F_j,F_global)/Ï )· G_Ï(i), where simâ(â ,â )sim(·,·) denotes cosine similarity, Ï is a temperature coefficient, and GÏâ(i)G_Ï(i) is a Gaussian window weight defined as GÏâ(i)=expâĄ(âdi2/2âÏ2)G_Ï(i)= (-d_i^2/2Ï^2), with did_i being the normalized distance from the i-th window center to the image center. This Gaussian weighting smoothly decays the contribution of windows near the image boundary, effectively suppressing stitching artifacts while preserving the semantic fidelity of central regions. The bandwidth Ï is set to 0.5 in our experiments to balance boundary smoothness and content preservation. This mechanism establishes implicit correlations among windows through the global anchor, ensuring that the output of each window is constrained not only by local visual information but also by the global contextual distribution. Meanwhile, the Gaussian weighting GÏâ(i)G_Ï(i) effectively eliminates prediction jumps at window seams, enabling seamless end-to-end inference on remote sensing images of arbitrary sizes and generalizing VFMs trained at limited resolutions to large-scale remote sensing scenarios. Experiments As a trainingâfree framework, DinoSplatâOV requires no preâtraining or parameter updates and performs endâtoâend inference directly at test time. Consequently, our experiments focus on performance evaluation and mainly compare against existing trainingâfree openâvocabulary segmentation methods. Datasets To comprehensively assess the generalisation capability of DinoSplatâOV across diverse remote sensing scenarios, we select four representative multiâcategory segmentation datasets: DOTA(Xia et al. 2018) (dense remote sensing objects), LoveDA (Wang et al. 2021)(ruralâurban mixed agricultural scenes), UDD5(Chen et al. 2018) (droneâstyle natural imagery), and ISPRS Vaihingen(Niemeyer et al. 2012)(urban scene without blue band). These datasets exhibit significant variations in spatial resolution, object density, and scene complexity, enabling a thorough evaluation of the modelâs adaptability to various remote sensing image types. Following SAMRS(Wang et al. 2023a) , we transform the original DOTA dataset into a semantic segmentation dataset with pixel-level annotations. Comparison with SOTA Methods Given that DinoSplatâOV is designed for trainingâfree inference, we select the most representative trainingâfree openâvocabulary segmentation models as our baselines. To ensure fair comparison, all competing methods are uniformly adapted with a slidingâwindow strategy to handle the large input sizes of remote sensing images, with window parameters kept consistent with those of DinoSplatâOV. Experimental results demonstrate that DinoSplatâOV achieves overall performance on par with the current stateâofâtheâart (SOTA), and exhibits particularly pronounced advantages in dense object scenarios (e.g., the DOTA dataset), validating the effectiveness of our approach in remoteâsensingâspecific segmentation tasks. The experimental results are presented in the figure above. DinoSplatâOV is built upon DINO.txt, while SegEarthâOV is based on ClearCLIP. It can be observed that ClearCLIP, by optimizing the CLIP architecture, achieves stronger imageâtext matching capability compared to DINO.txt, whereas DINO exhibits superior semantic clustering performance in the visual domain. Building upon this foundation, SegEarthâOV further introduces a preâtrained FeatUp upsampling module to attain stateâofâtheâart performance. In parallel, our approach incorporates Laplacian propagation and a Gaussianâsplattingâinspired testâtime optimization (TTO) upsampling module. Following a similar optimization trajectory, our method achieves performance on par with the SOTA. GSUP vs Other Feature Reconstruction Models To validate the effectiveness of GSUP, we compare it against two representative generalâpurpose upsampling models: AnyUp, a generic upsampler that supports DINO features, and SatUp, a dedicated upsampler trained on a remote sensing subset that is also compatible with DINO features. Experimental results demonstrate that GSUP, without any preâtraining, achieves feature reconstruction quality comparable to preâtrained upsamplers through testâtime optimization (TTO) alone, showing strong competitiveness across multiple remote sensing scenarios. In our experiments, They reconstruct 72Ă72 DINOv3 patch features into 224Ă224 dense feature maps, where each patch token is modeled as an adaptive Gaussian primitive for pixel-level semantic feature recovery. Method Pre Trained mIoU Infer Time Bilinearbaseline No 35.6 0.02s UPLiFTCVPR â 40.6 0.1s AnyUpICLR â 43.8 0.3s SatUp â (RS) 43.2 0.1s GSUPOurs TTO 42.9 1.2s Table 2: Performance of different feature upsampling methods on the ImageNet dataset subset. Figure 5: Visualization of feature reconstruction across different upsampling methods. Table 2 show that GSUP achieves slightly lower accuracy than the preâtrained optimal model, yet the gap is acceptableâreflecting a tradeâoff between performance and efficiency. Its core strength lies in being datasetâagnostic and preâtrainingâfree: unlike AnyUp or SatUp, which learn parameters on large datasets, GSUP optimizes only Gaussian splatting parameters per test image (10 SGD iterations). This yields a good balance among deployment flexibility, theoretically maximal crossâdomain generalization, and cost, suitable for remote sensing with scarce annotations and high variability. Also note that FeatUp needs local CUDA compilation, AnyUp requires NATTEN, while GSUP needs no preâtraining, offering better portability with minimal modifications. Ablations To validate the effectiveness of each of the four proposed modules in adapting DINO.txt for trainingâfree remote sensing openâvocabulary segmentation, we conduct ablation studies on the UDD5 dataset. Figure 6: The features extracted by DINO.txt are first enhanced and denoised through Textâaware Laplacian Propagation (TLP), and then reconstructed into pixelâlevel highâresolution representations via RGBâguided Gaussian Splatting Upsampling (GSUP), ultimately yielding the final semantic segmentation predictions. Method Base mIoU Î w/o all (baseline) 32.4 0 w/ Sliding Window 36.1 +3.7 w/ Synonym Aggregation 37.7 +1.6 w/ Laplacian Propagation (TLP) 40.7 +3.0 w/ Gaussian Splatting (GSUP) 42.9 +2.2 Table 3: Ablation study of the proposed modules on the UDD5 dataset. Table 3 shows that sliding window and synonym aggregation provide foundational improvements, increasing the mIoU from 32.4% to 36.1% and 37.7%, respectively. These gains demonstrate their effectiveness in handling large-scale remote sensing images and alleviating the text-image matching ambiguity of DINO.txt. The proposed core modules, TLP and GSUP, further improve the performance to 40.7% and 42.9%, corresponding to absolute gains of +8.3% and +10.5% over the baseline, respectively. TLP performs text-guided diffusion to enhance semantic consistency, suppressing noisy predictions while preserving category boundaries. GSUP addresses the spatial resolution limitation of DINOv3 patch features by reconstructing high-resolution semantic representations through RGB-guided anisotropic Gaussian aggregation with test-time optimization. The complementary effects of semantic refinement and resolution recovery lead to the best performance when all modules are combined. Generalizability of different backbones Backbone Base mIoU Î DINO.txt 37.7 - + TLP 40.7 +3.0 + GSUP 42.9 +2.2 ClearCLIP 38.2 - + TLP 40.8 +2.6 + GSUP 42.8 +2.0 Table 4: Replace different backbones on the UDD5 dataset. As shown in Table 4, we further investigate the generalizability of TLP and GSUP. While SegEarth-OV incorporates a featUp upsampling module pretrained on Million-AID(Long et al. 2021) dataset, which theoretically endows it with appreciable generalization capability, both TLP and GSUP require no dataset-specific pretraining at all. This allows them to achieve maximal generalization, effectively enabling Segment-Anything performance, particularly on less commonly used datasets. Conclusion DinoSplatâOV is the first trainingâfree remote sensing openâvocabulary segmentation framework built on the DINOv3 text encoder (DINO.txt). It requires no training or fineâtuning, relying on two core inferenceâtime modules: Textâaware Graph Laplacian Propagation (TLP) to align coarse features with semantic priors, and 2D Gaussian Splatting Upsampling (GSUP) to reconstruct lowâresolution features into highâprecision pixelâlevel predictions. Extensive experiments on UDD5, DOTA, LoveDA, Vaihingen and other benchmarks show that DinoSplatâOV achieves competitive or even superior performance over stateâofâtheâart trainingâfree methods, filling the gap of DINOâseries models in this task. Moreover, by transferring the explicit scene representation of 3D Gaussian splatting to 2D feature recovery for dense segmentation, this work offers a lightweight, preâtrainingâfree alternative for highâresolution feature reconstruction from frozen backbones, with generality extendable to other architectures like CLIP or SigLIP. References Q. Cao, Y. Chen, C. Ma, and X. Yang (2025) Open-vocabulary high-resolution remote sensing image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing. Cited by: Remote Sensing Adaptations and Domain Gaps.. L. Chambon, P. Couairon, E. Zablocki, A. Boulch, N. Thome, and M. Cord (2025) NAF: zero-shot feature upsampling via neighborhood attention filtering. External Links: Link Cited by: GSUP: Gaussian Splatting Inspired Upsampling. Y. Chen, Y. Wang, P. Lu, Y. Chen, and G. Wang (2018) Large-scale structure from motion with semantic constraints of aerial images. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), Cited by: Datasets. Y. Chen, X. Chen, A. Chen, G. Pons-Moll, and Y. Xiu (2025) Feat2gs: probing visual foundation models with gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 6348â6361. Cited by: 3D Gaussian Splatting. S. Cho, H. Shin, S. Hong, A. Arnab, P. H. Seo, and S. Kim (2024) CAT-seg: cost aggregation for open-vocabulary semantic segmentation. External Links: 2303.11797 Cited by: Remote Sensing Adaptations and Domain Gaps.. S. Fu, M. Hamilton, L. E. Brandt, A. Feldmann, Z. Zhang, and W. T. Freeman (2024) FeatUp: a model-agnostic framework for features at any resolution. In The Twelfth International Conference on Learning Representations, External Links: Link Cited by: Remote Sensing Adaptations and Domain Gaps.. C. Jose, T. Moutakanni, D. Kang, F. Baldassarre, T. Darcet, H. Xu, D. Li, M. Szafraniec, M. Ramamonjisoa, M. Oquab, O. SimĂ©oni, H. V. Vo, P. Labatut, and P. Bojanowski (2024) DINOv2 meets text: a unified framework for image- and pixel-level vision-language alignment. arXiv preprint arXiv:2412.16334. External Links: 2412.16334, Document Cited by: DINOv3 Text Encoder (DINO.txt). B. Kerbl, G. Kopanas, T. LeimkĂŒhler, and G. Drettakis (2023) 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42 (4). External Links: Link Cited by: 3D Gaussian Splatting. M. Lan, C. Chen, Y. Ke, X. Wang, L. Feng, and W. Zhang (2024a) Clearclip: decomposing clip representations for dense vision-language inference. In European Conference on Computer Vision, p. 143â160. Cited by: Training-Free Open-Vocabulary Segmentation Paradigms.. M. Lan, C. Chen, Y. Ke, X. Wang, L. Feng, and W. Zhang (2024b) Proxyclip: proxy attention improves clip for open-vocabulary segmentation. In European Conference on Computer Vision, p. 70â88. Cited by: Training-Free Open-Vocabulary Segmentation Paradigms.. B. Lee, H. S. Seong, S. Hyun, G. Park, W. Moon, and J. Heo (2026) Looking beyond the window: global-local aligned clip for training-free open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: Remote Sensing Adaptations and Domain Gaps.. K. Li, R. Liu, X. Cao, X. Bai, F. Zhou, D. Meng, and Z. Wang (2025) Segearth-ov: towards training-free open-vocabulary segmentation for remote sensing images. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 10545â10556. Cited by: Remote Sensing Adaptations and Domain Gaps.. Y. Long, G. Xia, S. Li, W. Yang, M. Y. Yang, X. X. Zhu, L. Zhang, and D. Li (2021) On creating benchmark dataset for aerial image interpretation: reviews, guidances and million-aid. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, p. 4205â4230. Cited by: Generalizability of different backbones. J. Niemeyer, F. Rottensteiner, and U. Soergel (2012) ISPRS 2d semantic labeling contest. ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences I-3, p. 293â298. Cited by: Datasets. V. StojniÄ, Y. Kalantidis, J. Matas, and G. Tolias (2025) LPOSS: label propagation over patches and pixels for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: Training-Free Open-Vocabulary Segmentation Paradigms.. D. Wang, J. Zhang, B. Du, M. Xu, L. Liu, D. Tao, and L. Zhang (2023a) SAMRS: scaling-up remote sensing segmentation dataset with segment anything model. In Advances in Neural Information Processing Systems, Vol. 36, p. 8815â8827. Cited by: Datasets. F. Wang, J. Mei, and A. Yuille (2023b) SCLIP: rethinking self-attention for dense vision-language inference. arXiv preprint arXiv:2312.01597. Cited by: Training-Free Open-Vocabulary Segmentation Paradigms.. J. Wang, Z. Zheng, A. Ma, X. Lu, and Y. Zhong (2021) LoveDA: a remote sensing land-cover dataset for domain adaptive semantic segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, J. Vanschoren and S. Yeung (Eds.), Vol. 1, p. . External Links: Link Cited by: Datasets. T. Wimmer, P. Truong, M. Rakotosaona, M. Oechsle, F. Tombari, B. Schiele, and J. E. Lenssen (2026) AnyUp: universal feature upsampling. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: GSUP: Gaussian Splatting Inspired Upsampling. M. WysoczaĆska, O. SimĂ©oni, M. Ramamonjisoa, A. Bursuc, T. TrzciĆski, and P. PĂ©rez (2024) CLIP-dinoiser: teaching clip a few dino tricks for open-vocabulary semantic segmentation. ECCV. Cited by: Training-Free Open-Vocabulary Segmentation Paradigms.. G. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang (2018) DOTA: a large-scale dataset for object detection in aerial images. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: Datasets.