Paper deep dive
SimFuse3D: Source-Guided Target Simulation and Confidence-Guided Multi-Stage Localization Reweighting for Cross-Platform 3D Object Detection
Yongchun Lin, Xinliang Zhang, Yun Zou, Zhixuan Xiao, Liang Lei, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, HaiKuo Xu, Haoang Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/7/2026, 6:59:11 AM
Summary
The paper introduces SimFuse3D, a method for cross-platform LiDAR unsupervised domain adaptation (UDA) that addresses 'box-point inconsistency' where pseudo-object boxes contain sparse or inconsistent point clouds. SimFuse3D uses an Object Memory to retrieve compatible labeled source instances and Target Simulation to align and replace the target's point observations with source geometry. It also employs Confidence-Guided Multi-Stage Localization Reweighting (CMLR) to adjust supervision weights. The method outperforms existing baselines like Pi3DET-Net across six cross-platform transfers.
Entities (9)
Relation Signals (8)
SimFuse3D → addresses → box-point inconsistency
confidence 95% · SimFuse3D... repairs the associated pseudo-object using measured geometry... We refer to this mismatch as box-point inconsistency.
SimFuse3D → outperforms → Pi3DET-Net
confidence 95% · Across six cross-platform transfers, SimFuse3D exceeds Pi3DET-Net on every reported AP metric
SimFuse3D → uses → Target Simulation
confidence 95% · Target Simulation places its ground-truth box at the target location, aligns its points with the target viewing geometry
SimFuse3D → uses → CMLR
confidence 95% · Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps each target pseudo-object confidence score to a bounded weight
SimFuse3D → uses → Object Memory
confidence 95% · Object Memory retrieves a compatible labeled source instance.
SimFuse3D → evaluatedon → nuScenes
confidence 90% · On nuScenes-to-KITTI, it ranks first
SimFuse3D → evaluatedon → KITTI
confidence 90% · On nuScenes-to-KITTI, it ranks first
Object Memory → uses → Point-NN
confidence 90% · We use the fixed, non-parametric Point-NN encoder... to obtain a retrieval key
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Changes in sensor height and viewpoint alter object-level point distributions, making cross-platform LiDAR unsupervised domain adaptation (UDA) difficult. Self-training uses labeled source scans and unlabeled target scans, yet a retained prediction may provide a useful target location while enclosing sparse foreground returns, background clutter, or points inconsistent with the predicted box. We refer to this mismatch as box-point inconsistency. We introduce SimFuse3D, which preserves the target placement and repairs the associated pseudo-object using measured geometry from labeled source scans. Object Memory retrieves a compatible labeled source instance. Target Simulation places its ground-truth box at the target location, aligns its points with the target viewing geometry, and filters the aligned crop to approximate the target observation. Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps each target pseudo-object confidence score to a bounded weight shared by RPN localization and R-CNN box regression. All components operate only during adaptation, leaving the detector architecture and inference graph unchanged. Across six cross-platform transfers, SimFuse3D exceeds Pi3DET-Net on every reported AP metric and ranks first among the compared adaptation methods on nearly all metrics. On nuScenes-to-KITTI, it ranks first among the compared adaptation methods with both evaluated detectors.
Tags
Links
- Source: https://arxiv.org/abs/2609.04886v1
- Canonical: https://arxiv.org/abs/2609.04886v1
Trouble viewing inline? Open PDF directly →
Full Text
46,420 characters extracted from source content.
Expand or collapse full text
SimFuse3D: Source-Guided Target Simulation and Confidence-Guided Multi-Stage Localization Reweighting for Cross-Platform 3D Object Detection Yongchun Lin Affiliation: Yongchun Lin, Yun Zou, Zhixuan Xiao, and Liang Lei are with Guangdong University of Technology. linyongchun,zouyun1, xiaozhixuan@mails.gdut.edu.cn; leiliang@gdut.edu.cn Xinliang Zhang Affiliation: Xinliang Zhang, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, and Haikuo Xu are with SenseTime. zhangxinliang, guojianya, zhaiyuqiang, wangxiaofeng, xuhaikuo@senseauto.com Yun Zou Affiliation: Yongchun Lin, Yun Zou, Zhixuan Xiao, and Liang Lei are with Guangdong University of Technology. linyongchun,zouyun1, xiaozhixuan@mails.gdut.edu.cn; leiliang@gdut.edu.cn Zhixuan Xiao Affiliation: Yongchun Lin, Yun Zou, Zhixuan Xiao, and Liang Lei are with Guangdong University of Technology. linyongchun,zouyun1, xiaozhixuan@mails.gdut.edu.cn; leiliang@gdut.edu.cn Liang Lei Affiliation: Yongchun Lin, Yun Zou, Zhixuan Xiao, and Liang Lei are with Guangdong University of Technology. linyongchun,zouyun1, xiaozhixuan@mails.gdut.edu.cn; leiliang@gdut.edu.cn Jianya Guo Affiliation: Xinliang Zhang, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, and Haikuo Xu are with SenseTime. zhangxinliang, guojianya, zhaiyuqiang, wangxiaofeng, xuhaikuo@senseauto.com Yuqiang Zhai Affiliation: Xinliang Zhang, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, and Haikuo Xu are with SenseTime. zhangxinliang, guojianya, zhaiyuqiang, wangxiaofeng, xuhaikuo@senseauto.com Xiaofeng Wang Affiliation: Xinliang Zhang, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, and Haikuo Xu are with SenseTime. zhangxinliang, guojianya, zhaiyuqiang, wangxiaofeng, xuhaikuo@senseauto.com HaiKuo Xu Affiliation: Xinliang Zhang, Jianya Guo, Yuqiang Zhai, Xiaofeng Wang, and Haikuo Xu are with SenseTime. zhangxinliang, guojianya, zhaiyuqiang, wangxiaofeng, xuhaikuo@senseauto.com Haoang Li †thanks: This work was supported by the Natural Science Foundation of Guangdong Province, China, under Grant 2023A1515010206.†thanks: *Corresponding authors: Liang Lei and Haoang Li. Affiliation: Haoang Li is with The Hong Kong University of Science and Technology (Guangzhou). haoangli@hkust-gz.edu.cn Abstract Changes in sensor height and viewpoint alter object-level point distributions, making cross-platform LiDAR unsupervised domain adaptation (UDA) difficult. Self-training uses labeled source scans and unlabeled target scans, yet a retained prediction may provide a useful target location while enclosing sparse foreground returns, background clutter, or points inconsistent with the predicted box. We refer to this mismatch as box–point inconsistency. We introduce SimFuse3D, which preserves the target placement and repairs the associated pseudo object using measured geometry from labeled source scans. Object Memory retrieves a compatible labeled source instance. Target Simulation places its ground-truth box at the target location, aligns its points with the target viewing geometry, and filters the aligned crop to approximate the target observation. Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps each target pseudo-object confidence score to a bounded weight shared by RPN localization and R-CNN box regression. All components operate only during adaptation, leaving the detector architecture and inference graph unchanged. Across six cross-platform transfers, SimFuse3D exceeds Pi3DET-Net on every reported AP metric and ranks first among the compared adaptation methods on nearly all metrics. On nuScenes→ , it ranks first among the compared adaptation methods with both evaluated detectors. I Introduction LiDAR-based 3D object detection supports autonomous vehicles and mobile robots. Most public benchmarks, including KITTI [1] and nuScenes [2], are collected from road vehicles. As documented by Pi3DET [3], a detector transferred to a drone or quadruped robot encounters a different sensor height, motion pattern, and target distribution. More importantly for point-based training, the new viewpoint changes incidence angles, self-occlusion, and sampling density within each object. Collecting accurate 3D boxes for every platform remains expensive. (a) Pi3DET-Net baseline (b) SimFuse3D adaptation (c) Vehicle→ detection Fig. 1: Motivation and qualitative comparison. (a) Pi3DET-Net [3] retains the target point observations enclosed by accepted pseudo boxes, allowing unreliable geometry to enter localization supervision. (b) SimFuse3D retrieves measured source geometry, aligns it with the target placement and viewpoint, and reweights target localization supervision with CMLR. (c) Vehicle→ example. The upper dashed green circle marks one ground-truth vehicle, and the lower circle marks two adjacent vehicles. All three boxes contain only a few LiDAR returns. Pi3DET-Net misses the three vehicles in the two marked regions, whereas SimFuse3D detects them. Red boxes show ground truth, blue boxes show predictions, and purple points are returns inside the ground-truth boxes. Unsupervised domain adaptation (UDA) offers a practical alternative to annotating each target platform: it combines labeled source scans with unlabeled target scans. In the cross-platform setting introduced by Pi3DET [3], we use Pi3DET-Net as the self-training baseline. Methods such as ST3D [4], ST3D++ [5], and MS3D++ [6] improve target supervision by filtering predicted boxes or refining them over time. Yet the points enclosed by an accepted box are usually retained as observed. Across platforms, this can leave a usable box paired with very few foreground returns, points concentrated on one side, or substantial background clutter. We refer to this mismatch as box–point inconsistency. It can corrupt localization and RoI supervision even when the pseudo-box location remains useful, and confidence filtering alone cannot correct the underlying observation. SimFuse3D constructs its retrieval memory directly from the labeled source scans already available in UDA. Figure 1(a)–(b) contrasts the unchanged-point update in Pi3DET-Net with our source-guided geometry repair and confidence weighting; the corresponding cross-platform results are reported in Tables I and I. SimFuse3D treats pseudo-box placement and the enclosed point observation separately. The target prediction supplies the placement; a retrieved source instance contributes measured foreground geometry and a ground-truth box. Object Memory selects this instance by similarity consensus. Target Simulation aligns it with the target position and viewing direction, filters the aligned crop by target-view angular limits, and replaces the inconsistent in-box observation. Object dimensions come from the retrieved source annotation rather than target-domain size statistics [7]. If no suitable match is found, the original pseudo object is kept. CMLR maps every target pseudo-object score to a bounded weight shared by RPN localization and R-CNN box regression. Simulated objects receive unit weight, whereas unmatched objects are reweighted from their retained teacher scores. These operations are confined to adaptation; the detector and its inference graph are unchanged (Fig. 2). This paper makes three contributions: • We identify box–point inconsistency in cross-platform self-training, where a usable pseudo-box location is paired with sparse, contaminated, or geometrically inconsistent points; confidence and point support alone do not fully indicate pseudo-object quality. • We propose SimFuse3D to repair and reweight target pseudo objects. Object Memory and Target Simulation replace accepted observations with target-view-aligned source geometry, while CMLR applies bounded score-derived weights to all target pseudo objects at the RPN and R-CNN localization stages. The design changes neither the detector architecture nor its inference graph. • Across six transfers and two detectors, SimFuse3D improves all 48 AP components over Pi3DET-Net and ranks first among the listed adaptation methods in 47 of the 48 comparisons; adding CMLR improves all eight TS-only ablation metrics. Fig. 2: SimFuse3D overview. Object Memory encodes labeled source objects for target pseudo-object retrieval. Target Simulation aligns the retrieved geometry with the target placement and viewpoint. Confidence-Guided Multi-Stage Localization Reweighting (CMLR) maps all target pseudo-object scores to bounded weights for RPN localization and R-CNN box regression, leaving the other losses unchanged. All three components are absent at inference. I Related Work I-A Unsupervised Domain Adaptation for 3D Detection Prior LiDAR adaptation methods reduce domain discrepancy at several levels. Statistical normalization [7] corrects object-size bias, SPG [8] generates semantic foreground points, and GPA-3D [9] aligns object features through geometry-aware prototypes. Sensor changes motivate a different set of techniques: CL3D [10] aligns spatial geometry and temporal motion across configurations, whereas DTS [11] combines beam resampling with cross-density consistency. These methods operate on size statistics, representations, or scan density rather than explicitly repairing the box-interior observation of a retained target pseudo object. Target-supervision methods address a complementary part of the problem. ST3D [4] maintains pseudo labels over time, and ST3D++ [5] strengthens their denoising and training. ReDB [12] selects pseudo labels according to reliability, diversity, and class balance. MS3D++ [6] combines multiple source experts with temporal refinement, although it assumes a multi-source setting. Object-level processing is more closely related to our work: PERE [13] removes points from unreliable boxes or substitutes high-confidence target instances, DALI [14] synthesizes points from prepared object-model and sensor libraries, and DiffRefine [15] uses diffusion to densify sparse proposals before second-stage refinement. In SimFuse3D, replacement geometry is retrieved directly from the labeled source scans and aligned with the target placement and viewing direction. The target pseudo box determines the insertion location, while the retrieved source instance provides the measured points and ground-truth box dimensions; no separate CAD or sensor-model library is needed. The information available during adaptation also differs across methods. Attentive Prototypes [16] assumes that source data are no longer accessible. CMDA [17] uses synchronized images as a semantic bridge, while MMAssist [18] combines image and text features and augments pseudo labels with an external 2D detector. These settings are relevant to cross-domain detection, but they are not equivalent to the single-source, LiDAR-only protocol used by SimFuse3D. I-B Point Cloud Representation and Cross-Platform Adaptation PointNet [19] and PointNet++ [20] introduced global and hierarchical point-set representations. Point-N [21] later showed that sampling, neighborhood grouping, encoding, and pooling also yield useful non-parametric descriptors. We use this descriptor only to retrieve source instances; it is not part of the detector. Cross-platform evaluation is supported by M3ED [22], which contains synchronized vehicle, quadruped, and drone data, and by Pi3DET [3], which provides the corresponding 3D detection benchmarks. UADA3D [23] studies a related setting with sparse LiDAR and large platform gaps. Our focus is narrower: the mismatch between a target pseudo box and the points used to supervise it. I Method I-A Problem Formulation and Overview Let s=(Psn,Ysn)n=1NsD_s=\(P_s^n,Y_s^n)\_n=1^N_s denote a labeled source domain, where PsnP_s^n is a point cloud and YsnY_s^n contains the corresponding 3D ground-truth boxes. Let t=Ptmm=1NtD_t=\P_t^m\_m=1^N_t denote an unlabeled target domain collected by a different sensing platform or sampled from a different dataset. Following the self-training protocol of Pi3DET [3], the detector periodically generates pseudo labels for target scan PtmP_t^m: Y^tm=(b^i,y^i,si)i=1Nm, Y_t^m=\( b_i, y_i,s_i)\_i=1^N_m, (1) where NmN_m is the number of retained predictions. Each pseudo label consists of a box b^i=(xi,yi,zi,li,wi,hi,θi)∈ℝ7 b_i=(x_i,y_i,z_i,l_i,w_i,h_i, _i) ^7, a class label y^i y_i, and a detection score si∈[0,1]s_i∈[0,1]. We use Oit=p∈Ptm∣p∈b^iO_i^t=\p∈ P_t^m p∈ b_i\ to denote the target points enclosed by b^i b_i. In the target domains considered here, some retained boxes contain only a few points; the observed points may cluster on one side of the box, and their outline may not fit the predicted dimensions. We use box–point inconsistency to describe these cases. This concern is particularly relevant early in self-training, when target predictions generally have lower scores. A predicted box can therefore provide a useful location cue without providing an equally reliable object observation for training. At each scheduled pseudo-label refresh, SimFuse3D first queries Object Memory with OitO_i^t. If the retrieval gate accepts a source object, Target Simulation removes the original points inside b^i b_i and inserts the matched source observation after aligning it with the target placement and viewing geometry. This produces a modified scan P~tm P_t^m and pseudo labels Y~tm Y_t^m, pairing the usable target location with measured source points and a source ground-truth box. If retrieval fails, we retain the original pseudo label. A missing match only indicates that the source memory does not contain a suitable instance, not that the target prediction is incorrect; the original pseudo label may still be correctly localized and provide useful supervision. During the subsequent detector update, CMLR converts pseudo-label scores into bounded object weights for RPN localization and R-CNN box regression. Operations outside these steps follow the Pi3DET-Net baseline. Fig. 3: Object Memory and Target Simulation. Object Memory stores source points, ground-truth boxes, metadata, and fixed Point-N [21] descriptors. A target pseudo object queries the memory for a labeled source instance. Target Simulation reloads the measured source points, retains the source dimensions while aligning the box with the target placement, and applies target-view alignment and filtering. I-B Object Memory Memory construction. Object Memory is constructed offline from the labeled source training scans. For each source ground-truth box bjsb_j^s, we record its class yjsy_j^s and the corresponding local LiDAR observation OjsO_j^s. Both memory entries and target queries use a fixed input of 256 local points. A crop containing more than 256 points is reduced by farthest-point sampling (FPS). For a sparse crop, we enlarge the box by a fixed metric margin to collect nearby measurements before sampling. The xyz coordinates are then translated by the box center but are not rescaled by the box dimensions. We use the fixed, non-parametric Point-N encoder [21], denoted by g(⋅)g(·), to obtain a retrieval key: zjs z_j^s =g(Ojs,bjs)∥g(Ojs,bjs)∥2, = g(O_j^s,b_j^s) g(O_j^s,b_j^s) _2, (2) ℳ =(zjs,bjs,Ojs,yjs,mj)j=1Nℳ. =\(z_j^s,b_j^s,O_j^s,y_j^s,m_j)\_j=1^N_M. Here, zjsz_j^s is a 2,304-dimensional unit vector used only for retrieval. It is computed from raw point coordinates and is not concatenated with detector backbone features. Thus, Object Memory does not require a separately trained retrieval network when the detector is updated. Each entry also stores the source ground-truth box, point record, class label, frame and box indices, and the source LiDAR origin in mjm_j. Target Simulation uses these records to reload the points inside the source box and recover their sensor-relative viewing direction. The memory contains source instances belonging to the categories used by the target training task. The construction supports multiple categories, although the experiments in this paper use a single car category. The memory is built once and remains fixed throughout target-domain adaptation. Gated retrieval. At each scheduled pseudo-label update, the points associated with a retained target candidate are centered at its predicted box and encoded in the same way to form a unit query qiq_i. In our implementation, an HNSW index produces the initial candidate set, after which the candidates are re-ranked by cosine similarity aij=qi⊤zjsa_ij=q_i z_j^s. Let πi(k) _i(k) denote the index of the k-th nearest entry, and let iN_i contain the top K valid entries whose similarity is at least δn _n. Candidate i is associated with a source object only if aiπi(1) a_i _i(1) ≥δ1, ≥ _1, (3) vote(yjs∣j∈i) \! (\y_j^s j _i\ ) >0. >0. The two conditions require a sufficiently similar nearest source instance and a positive majority vote among the qualified neighbors. When both hold, the top-ranked memory entry is passed to Target Simulation and the voted class is assigned to the candidate. Because the inserted object points and replacement box both come from a labeled source instance, we set the replacement confidence to 1.01.0. Retrieval is based on Point-N similarity and neighbor voting; no additional score shield or box-IoU gate is used. If the gate fails, the original positive pseudo label and its teacher score are retained. A failed match means that the current memory does not contain a suitable source instance, rather than that the target pseudo label is necessarily incorrect, so it should not be removed merely because retrieval fails. I-C Target Simulation TABLE I: Cross-platform adaptation to Pi3DET Quadruped and Drone. Entries are APBEV/AP3DAP_BEV/AP_3D (%). Entries within each detector column share the detector and evaluation configuration; Pi3DET-Net is reproduced under this setup. ST3D/ST3D++ use random object scaling unless ‡ marks no ROS. “Target Platform” is fully supervised; the best scores among the listed adaptation methods are bolded. Source Domain Method Pi3DET (Quadruped) Pi3DET (Drone) PV-RCNN Voxel R-CNN PV-RCNN Voxel R-CNN AP@0.7 AP@0.5 AP@0.7 AP@0.5 AP@0.7 AP@0.5 AP@0.7 AP@0.5 nuScenes Source Only 40.11 / 31.23 43.95 / 42.00 41.84 / 33.39 45.27 / 43.36 35.56 / 24.96 38.36 / 35.76 37.84 / 24.21 44.53 / 39.60 ST3D [4] 54.83 / 42.66 58.94 / 56.80 50.91 / 41.38 53.25 / 51.69 51.56 / 31.54 58.35 / 51.54 52.20 / 33.53 56.85 / 51.29 ST3D‡ [4] 54.68 / 43.34 58.64 / 57.78 50.46 / 40.95 53.02 / 51.20 50.94 / 31.84 57.68 / 51.07 51.95 / 33.25 56.65 / 51.14 ST3D++ [5] 56.08 / 41.90 58.84 / 58.03 51.05 / 41.45 53.76 / 51.70 51.26 / 33.35 56.63 / 51.34 54.32 / 34.70 59.41 / 54.03 ST3D++‡ [5] 52.48 / 40.41 56.56 / 54.17 50.90 / 41.42 53.12 / 50.97 53.03 / 34.18 58.45 / 53.02 52.97 / 34.79 58.54 / 52.52 ReDB [12] 52.82 / 41.80 56.16 / 54.44 51.67 / 40.21 54.72 / 52.84 52.34 / 34.24 57.95 / 52.65 50.85 / 34.69 54.63 / 49.87 MS3D++ [6] 57.36 / 45.28 61.02 / 59.08 52.94 / 43.79 55.68 / 55.16 55.59 / 33.78 61.20 / 55.57 55.02 / 36.41 59.95 / 54.35 Pi3DET-Net [3] 56.60 / 45.57 61.37 / 59.45 54.31 / 43.99 57.15 / 55.25 55.30 / 37.37 62.30 / 55.13 55.72 / 35.95 60.22 / 55.15 Ours 58.19 / 47.15 63.02 / 61.08 57.31 / 46.49 61.43 / 59.63 57.32 / 37.81 62.82 / 56.82 56.68 / 38.54 62.18 / 56.24 Target Platform 54.15 / 40.24 67.67 / 46.11 58.63 / 54.96 70.04 / 66.14 54.90 / 39.74 68.52 / 46.53 56.46 / 55.19 70.67 / 61.42 Pi3DET (Vehicle) Source Only 47.57 / 36.73 50.99 / 49.04 50.34 / 37.74 53.43 / 52.37 52.76 / 32.55 56.13 / 50.74 50.12 / 33.46 55.12 / 49.86 ST3D [4] 54.70 / 42.22 58.92 / 56.71 50.68 / 41.24 53.33 / 51.05 58.12 / 34.95 63.10 / 57.14 55.38 / 37.62 58.61 / 55.38 ST3D‡ [4] 54.88 / 42.85 59.23 / 56.40 51.14 / 40.88 55.58 / 53.27 56.58 / 34.58 63.20 / 56.02 53.33 / 37.57 56.42 / 53.29 ST3D++ [5] 54.84 / 42.56 58.15 / 56.26 51.06 / 41.94 53.56 / 52.96 57.13 / 34.71 61.93 / 56.23 55.56 / 39.25 60.96 / 55.64 ST3D++‡ [5] 54.82 / 43.14 58.62 / 56.51 54.78 / 43.45 57.48 / 55.22 57.16 / 34.61 61.94 / 56.15 55.51 / 39.47 58.70 / 55.64 ReDB [12] 53.68 / 41.84 56.87 / 54.80 48.47 / 36.86 51.71 / 49.76 61.93 / 38.31 67.04 / 60.93 54.76 / 36.10 59.72 / 53.98 MS3D++ [6] 55.88 / 44.07 60.26 / 57.54 50.87 / 39.91 55.30 / 53.08 64.99 / 39.69 69.64 / 63.57 60.02 / 41.20 63.40 / 59.98 Pi3DET-Net [3] 57.22 / 47.17 63.69 / 61.34 58.76 / 47.44 63.68 / 61.22 62.64 / 44.46 67.69 / 63.56 65.82 / 49.59 71.22 / 65.57 Ours 60.09 / 48.79 64.01 / 62.02 62.54 / 49.34 67.71 / 65.40 64.39 / 45.86 69.65 / 63.98 67.58 / 50.77 73.05 / 69.35 Target Platform 54.15 / 40.24 67.67 / 46.11 58.63 / 54.96 70.04 / 66.14 54.90 / 39.74 68.52 / 46.53 56.46 / 55.19 70.67 / 61.42 Target Simulation is applied only when Object Memory returns a valid source instance (Ojs,bjs)(O_j^s,b_j^s) for target candidate (Oit,b^i)(O_i^t, b_i). As detailed in Fig. 3, the target box supplies the placement, while the source ground truth supplies the object dimensions and measured point structure. We do not scale the source points to the dimensions of the target pseudo box. Pose alignment. Let oso_s and oto_t be the source and target LiDAR origins. We use the coordinate-wise median of each object crop, denoted by μs _s and μt _t, to form the sensor-to-object rays vs=μs−osv_s= _s-o_s and vt=μt−otv_t= _t-o_t. The source crop and its ground-truth box are rotated together by the difference between the two ray azimuths: Δψi=atan2(vt,y,vt,x)−atan2(vs,y,vs,x). _i=atan2(v_t,y,v_t,x)-atan2(v_s,y,v_s,x). (4) After the yaw rotation, the box center is translated so that its bird’s-eye-view center matches (xi,yi)(x_i,y_i) and its bottom face matches that of b^i b_i. Since the source height is retained, the resulting reference center is c¯i=(xi,yi,zi−hi/2+hjs/2) c_i=(x_i,y_i,z_i-h_i/2+h_j^s/2). For p∈Ojsp∈ O_j^s, the aligned point is p~=c¯i+Rz(θjs+Δψi)Rx(Δϕi)Rz(−θjs)(p−cjs), p= c_i+R_z( _j^s+ _i)R_x( _i)R_z(- _j^s)(p-c_j^s), (5) where cjsc_j^s and θjs _j^s are the center and yaw of the source ground-truth box. For cross-platform pairs, Δϕi _i is obtained from the difference between the median source and target elevation angles and clipped by a task-specific bound ϕmax _ . This pitch correction is applied in the source-box coordinate system and changes only the points. The replacement box keeps the source dimensions and has horizontal heading θjs+Δψi _j^s+ _i. Point distribution simulation and replacement. After pose alignment, we filter the transformed source crop from the target LiDAR view. Let i=p~∣p∈OjsA_i=\ p p∈ O_j^s\ be the aligned crop. For q∈iq _i, α(q)α(q) and ε(q) (q) denote the absolute azimuth and elevation deviations between the point ray q−otq-o_t and the center ray c¯i−ot c_i-o_t. We retain i=q∈i∣α(q)≤τaz,ε(q)≤τel,V_i=\q _i α(q)≤ _az,\ (q)≤ _el\, where τaz _az and τel _el are angular limits. If |i|<nmin|V_i|<n_ , we retain iA_i to avoid replacing the target observation with an overly sparse crop. The pseudo box specifies the placement but does not rescale the source crop. Object dimensions and local point geometry come from the retrieved source instance. The angular filter selects the visible subset under the target view. Point coordinates remain unchanged after filtering, and the point count is not matched to that of the target crop. Cropping the retained points to the transformed source box gives O~it O_i^t. We then replace the points inside b^i b_i: P~tm=(Ptm∖Oit)∪O~it. P_t^m=(P_t^m O_i^t)∪ O_i^t. (6) The associated pseudo label uses the translated and rotated source ground-truth box. Points outside the original target pseudo box are not erased, and unmatched pseudo objects are left unchanged. I-D Confidence-Guided Multi-Stage Localization Reweighting Target Simulation corrects matched pseudo objects, but unmatched predictions are retained and their quality can vary. Cross-platform domain shifts increase false positives and localization errors; treating all pseudo objects equally would therefore give unreliable boxes the same supervision strength as reliable ones. We therefore introduce CMLR, an object-level reliability mechanism that applies pseudo-object confidence consistently at the proposal and RoI localization stages. For every retained target pseudo object, CMLR first maps its score sis_i to ω~i=max(siγ,ωmin), ω_i= (s_i^γ, _ ), (7) ωi=βω~i+(1−β), _i=β ω_i+(1-β), (8) where γ controls the sensitivity to the confidence score, ωmin _ prevents low-score objects from losing all supervision, and β∈[0,1]β∈[0,1] balances confidence weighting with unit weighting. Thus, high-score pseudo objects provide stronger supervision, whereas low-score objects remain in training with reduced gradient contribution. Source objects and simulated replacements use unit weight; the latter have confidence reset to 1.01.0 after introducing source dimensions and measured geometry. The same object weight is applied to proposal-stage localization and R-CNN box regression. Let ℬ=RPN-loc,R-CNN-regB=\RPN -loc,R -CNN -reg\ and let ℒotherL_other contain the classification terms and other unchanged detector and adaptation losses. The target objective is ℒours=ℒother+∑b∈ℬλb∑iωiℒb,imax(1,∑iωi),L_ours=L_other+ _b _b _i _iL_b,i (1, _i _i), (9) where λb _b is the original coefficient of branch b. The normalization keeps the loss scale stable across mini-batches with different numbers of pseudo objects. Source-domain ground-truth objects retain unit weight. Uncertain target predictions have less influence on optimization, and the detector architecture is unchanged. I-E Training and Inference Adaptation starts from the source-pretrained Pi3DET-Net detector. Object Memory is built once from labeled source objects. At each scheduled pseudo-label refresh, the current detector serves as the pseudo-label teacher and predicts boxes and confidence scores for the target training scans. The gate in Eq. (3) pairs eligible predictions with source instances, which are then processed by Target Simulation. If no match is accepted, the original pseudo box and its points are retained. The updated pseudo labels and replacement points are reused until the next refresh. TABLE I: Bidirectional adaptation between Pi3DET Quadruped (Q) and Drone (D). Entries are APBEV/AP3DAP_BEV/AP_3D (%). Entries within each detector column share the detector and evaluation configuration; Pi3DET-Net is reproduced under this setup. “Target Platform” is fully supervised; the best scores among the listed adaptation methods are bolded. Task Method PV-RCNN Voxel R-CNN AP@0.7 AP@0.5 AP@0.7 AP@0.5 Q→ Source Only 27.52 / 11.76 32.60 / 27.62 28.02 / 12.14 33.93 / 28.54 ST3D [4] 31.14 / 13.93 38.27 / 31.90 30.31 / 20.06 36.03 / 33.05 ST3D++ [5] 34.35 / 16.46 41.97 / 36.98 33.88 / 20.89 39.29 / 36.31 ReDB [12] 27.71 / 16.79 33.57 / 29.66 32.21 / 22.57 36.80 / 34.13 MS3D++ [6] 31.77 / 16.47 40.00 / 35.27 26.68 / 19.43 29.41 / 29.20 Pi3DET-Net [3] 40.62 / 26.87 47.60 / 44.34 38.04 / 20.37 43.74 / 40.42 Ours 41.20 / 28.57 48.22 / 44.89 39.89 / 23.22 45.63 / 42.33 Target Platform 54.90 / 39.74 68.52 / 46.53 56.46 / 55.19 70.67 / 61.42 D→ Source Only 39.18 / 26.46 44.90 / 40.80 35.79 / 24.85 39.99 / 37.37 ST3D [4] 50.04 / 39.16 54.73 / 52.36 46.01 / 32.18 51.95 / 48.14 ST3D++ [5] 51.80 / 40.94 56.29 / 53.88 51.15 / 35.62 54.30 / 51.82 ReDB [12] 46.32 / 34.36 50.51 / 48.23 46.57 / 34.31 53.03 / 50.43 MS3D++ [6] 50.56 / 38.46 55.50 / 53.13 49.10 / 35.09 52.40 / 50.28 Pi3DET-Net [3] 53.14 / 41.65 58.87 / 56.49 52.96 / 38.24 59.98 / 56.36 Ours 56.37 / 43.49 62.54 / 59.08 54.52 / 41.46 61.67 / 59.11 Target Platform 54.15 / 40.24 67.67 / 46.11 58.63 / 54.96 70.04 / 66.14 Source and target samples are optimized with the same detector. Source supervision follows the baseline, whereas Eq. (9) reweights the RPN localization and R-CNN regression losses of target pseudo objects; the classification losses are unchanged. At inference, only the adapted detector is used. Memory lookup, point replacement, and loss weighting are not invoked, so the inference graph is identical to that of the baseline. IV Experiments IV-A Experimental Setup Datasets and evaluation. For the six cross-platform transfers, we follow the unsupervised Pi3DET protocol [3]: source annotations are available for training, whereas target annotations are used only for evaluation. The transfers span nuScenes and the Pi3DET Vehicle, Quadruped (Q), and Drone (D) platforms; we also evaluate nuScenes→ . All experiments use the car category and report APBEVAP_BEV and AP3DAP_3D over 40 recall positions. Cross-platform results use IoU thresholds of 0.70.7 and 0.50.5, while nuScenes→ follows the standard car protocol at IoU 0.70.7. TABLE I: nuScenes→ adaptation with PV-RCNN at IoU 0.70.7. Entries are AP (%); the best adaptation scores are bolded. “w/ SN” denotes statistical normalization, and “Target Platform” is fully supervised. Method BEVAP_BEV 3DAP_3D Source Only 64.98 38.85 SN [7] 55.00 43.95 ST3D [4] 74.47 50.24 ST3D [4] w/ SN 78.40 70.90 ST3D++ [5] 81.70 45.35 ST3D++ [5] w/ SN 84.98 75.50 ReDB [12] 81.26 50.79 DTS [11] 83.90 71.80 PERE [13] 82.09 68.34 CMDA [17] 84.85 75.02 DALI [14] 84.12 75.43 Ours 85.94 76.16 Target Platform 88.98 82.50 TABLE IV: nuScenes→ adaptation with Voxel R-CNN at IoU 0.70.7. Entries are AP (%); the best adaptation scores are bolded. Unreported methods are omitted, and “Target Platform” is fully supervised. Method BEVAP_BEV 3DAP_3D Source Only 41.58 17.62 SN [7] 33.95 23.26 ST3D [4] 77.29 35.58 ST3D [4] w/ SN 87.11 66.02 ST3D++ [5] 82.08 38.10 ST3D++ [5] w/ SN 85.84 68.64 ReDB [12] 80.58 34.82 Ours 87.75 72.84 Target Platform 91.25 85.29 Implementation. We implement all experiments with PyTorch [24] and OpenPCDet [25], using PV-RCNN [26] and Voxel R-CNN [27] as in Pi3DET-Net [3]. Entries within each detector column share the same detector and evaluation settings; Pi3DET-Net results are our reproductions. Its official nuScenes→ result uses SECOND-IoU and is therefore not directly comparable to our PV-RCNN/Voxel R-CNN runs. Adaptation lasts 15 epochs with Adam [28] and a OneCycle schedule [29]; pseudo-label refresh intervals follow the corresponding baseline. Object Memory retrieves K=5K=5 neighbors, Target Simulation applies angular filtering, and CMLR weights RPN localization and R-CNN box regression. IV-B Comparison with Adaptation Methods Tables I and I compare source-only training and the listed adaptation methods on the six cross-platform transfers. Table I distinguishes runs with and without random object scaling (ROS), while Table I reports the direct Quadruped–Drone transfers. Tables I and IV report the nuScenes→ results separately for PV-RCNN and Voxel R-CNN, including the available statistical-normalization (SN) variants. Cross-platform adaptation. Across six transfers and two detectors, SimFuse3D exceeds Pi3DET-Net in all 48 AP components and ranks first among the listed adaptation methods in 47. The exception is PV-RCNN APBEV@0.7AP_BEV@0.7 on Vehicle→ , where MS3D++ obtains 64.9964.99 versus 64.3964.39. The largest margin is on nuScenes→ : Voxel R-CNN AP3D@0.5AP_3D@0.5 rises from 55.2555.25 to 59.6359.63 (+4.38+4.38 points). For the two Vehicle-source transfers, all 16 components improve. Figure 5 shows three Vehicle→ misses recovered over Pi3DET-Net at IoU 0.50.5. Cross-dataset adaptation. On nuScenes→ with PV-RCNN, SimFuse3D records 85.94/76.1685.94/76.16, compared with 84.98/75.5084.98/75.50 for ST3D++ w/ SN. The gains are 0.960.96 points in APBEVAP_BEV and 0.660.66 points in AP3DAP_3D. For Voxel R-CNN, the strongest adaptation-baseline values come from different rows: ST3D w/ SN reaches 87.1187.11 in APBEVAP_BEV, whereas ST3D++ w/ SN reaches 68.6468.64 in AP3DAP_3D. SimFuse3D obtains 87.75/72.8487.75/72.84, improving these values by 0.640.64 and 4.204.20 points, respectively. Efficiency. With Voxel R-CNN on Vehicle→ , SimFuse3D uses the same detector as Pi3DET-Net and has nearly identical optimization time per epoch. Extra computation is confined to two scheduled pseudo-label refreshes, where retrieval/voting and Target Simulation add about 1.01.0 min in total. Thus, 15-epoch adaptation takes 20.420.4 min versus 19.219.2 min for Pi3DET-Net. These operations are omitted at test time, preserving the detector and inference graph. TABLE V: Component ablation with Voxel R-CNN on the Vehicle-source transfers. TS and CMLR denote Target Simulation and Confidence-Guided Multi-Stage Localization Reweighting. Entries are APBEV/AP3DAP_BEV/AP_3D (%); best scores are bolded. TS CMLR Pi3DET (Quadruped) Pi3DET (Drone) AP@0.7 AP@0.5 AP@0.7 AP@0.5 58.76 / 47.44 63.68 / 61.22 65.82 / 49.59 71.22 / 65.57 ✓ 60.94 / 48.22 65.97 / 63.58 67.53 / 49.66 70.94 / 67.36 ✓ ✓ 62.54 / 49.34 67.71 / 65.40 67.58 / 50.77 73.05 / 69.35 TABLE VI: Voxel R-CNN recall by target ground-truth point support. Entries are Pi3DET-Net/Ours (Δ ), in percent; V, Q, and D denote Vehicle, Quadruped, and Drone. #GT is the object count, Dist. is the median sensor distance, and Δ is computed before rounding. Task GT pts. #GT Dist. (m) R@0.5 R@0.7 V→ 0–15 304 28.24 9.54/11.51 (+1.97) 4.93/5.26 (+0.33) 16–30 180 25.16 45.56/51.11 (+5.56) 27.22/30.00 (+2.78) ≥31≥ 31 1,581 18.46 80.83/86.08 (+5.25) 66.86/70.46 (+3.61) All 2,065 20.41 67.26/72.06 (+4.79) 54.29/57.34 (+3.05) V→ 0–15 1,573 36.10 29.37/36.62 (+7.25) 16.59/19.64 (+3.05) 16–30 871 28.77 63.15/65.21 (+2.07) 42.37/44.32 (+1.95) ≥31≥ 31 3,452 16.01 85.75/86.88 (+1.13) 74.02/74.10 (+0.09) All 5,896 20.36 67.37/70.27 (+2.90) 54.02/55.17 (+1.15) IV-C Ablation Study Table V reports the Voxel R-CNN ablation on the two Vehicle-source transfers. The second row adds Target Simulation (TS), and the last row further adds CMLR. Target Simulation alone improves seven of the eight metrics, with its largest gains on Quadruped at IoU 0.50.5; the only decrease is 0.280.28 points in Drone APBEV@0.5AP_BEV@0.5. Geometry repair acts only on pseudo objects with an accepted memory match, so unmatched objects still contribute their original observations and, without CMLR, receive the same localization weight. CMLR retains these potentially correct pseudo labels but reduces the influence of lower-confidence cases. The full configuration consequently improves all eight metrics over both the TS-only row and the baseline. Fig. 4: Pseudo-label reliability on Vehicle→ . Each cell reports Q@IoU≥0.5Q@IoU≥ 0.5 (%) and the sample count n for a teacher-confidence and in-box point-support bin. IV-D Pseudo-label Reliability and Point-Sparse Objects Reliability analysis. Figure 4 groups Vehicle→ pseudo boxes by teacher confidence and in-box point count. Here, Q is the fraction whose best ground-truth match reaches IoU ≥0.5≥ 0.5. Across confidence bins, Q rises from 76.2%76.2\% for objects with 0–15 points to 93.4%93.4\% for those with at least 31 points, although the trend is not monotonic in every interval. Confidence and point support are therefore complementary rather than sufficient quality indicators. A failed retrieval only means that no suitable source instance was found; the original pseudo label may still provide localization supervision. SimFuse3D retains unmatched pseudo labels, downweights their localization targets with CMLR, and assigns unit weight to simulated objects. Recall by target point support. Table VI stratifies final-detector recall by target point support, with positive gains in all 16 entries. The largest gain is 7.257.25 points at IoU 0.50.5 for the 0–15-point Vehicle→ bin. Overall Recall@0.5 increases by 4.794.79 points on Vehicle→ and 2.902.90 points on Vehicle→ . Gains at IoU 0.70.7 are smaller, indicating reduced benefit under stricter overlap. Fig. 5: Vehicle→ comparison. Orange outlines mark three sparse vehicles recovered by SimFuse3D at IoU 0.50.5 but missed by Pi3DET-Net. V Conclusion A useful pseudo-box location does not guarantee a reliable point observation. SimFuse3D preserves the former, repairs the latter with measured source geometry, and uses CMLR to control localization supervision from retained pseudo labels. Across six transfers, SimFuse3D improves all 48 AP components over Pi3DET-Net by up to 4.384.38 points and attains the highest listed nuScenes→ adaptation scores with both detectors. Memory lookup, Target Simulation, and CMLR operate only during adaptation, leaving the deployed detector unchanged. Fixed retrieval thresholds and a stored source-object memory are current limitations; adaptive retrieval and more compact memory construction remain future directions. References [1] A. Geiger, P. Lenz, and R. Urtasun (2012) Are we ready for autonomous driving? the KITTI vision benchmark suite. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 3354–3361. External Links: Document Cited by: §I. [2] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom (2020) nuScenes: a multimodal dataset for autonomous driving. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 11621–11631. Cited by: §I. [3] A. Liang, L. Kong, D. Lu, Y. Liu, J. Fang, H. Zhao, and W. T. Ooi (2025) Perspective-invariant 3D object detection. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), p. 27725–27738. Cited by: Fig. 1, §I, §I, §I-B, §I-A, TABLE I, TABLE I, TABLE I, TABLE I, §IV-A, §IV-A. [4] J. Yang, S. Shi, Z. Wang, H. Li, and X. Qi (2021) ST3D: self-training for unsupervised domain adaptation on 3D object detection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 10368–10378. Cited by: §I, §I-A, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE IV, TABLE IV. [5] J. Yang, S. Shi, Z. Wang, H. Li, and X. Qi (2023) ST3D++: denoised self-training for unsupervised domain adaptation on 3D object detection. IEEE Trans. Pattern Anal. Mach. Intell. 45 (5), p. 6354–6371. External Links: Document Cited by: §I, §I-A, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE IV, TABLE IV. [6] D. Tsai, J. S. Berrio, M. Shan, E. Nebot, and S. Worrall (2025) MS3D++: ensemble of experts for multi-source unsupervised domain adaptation in 3D object detection. IEEE Trans. Intell. Veh. 10 (3), p. 1999–2014. External Links: Document Cited by: §I, §I-A, TABLE I, TABLE I, TABLE I, TABLE I. [7] Y. Wang, X. Chen, Y. You, L. E. Li, B. Hariharan, M. Campbell, K. Q. Weinberger, and W. Chao (2020) Train in Germany, test in the USA: making 3D object detectors generalize. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 11713–11723. External Links: Document Cited by: §I, §I-A, TABLE I, TABLE IV. [8] Q. Xu, Y. Zhou, W. Wang, C. R. Qi, and D. Anguelov (2021) SPG: unsupervised domain adaptation for 3D object detection via semantic point generation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), p. 15446–15456. Cited by: §I-A. [9] Z. Li, J. Guo, T. Cao, B. Liu, and W. Yang (2023) GPA-3D: geometry-aware prototype alignment for unsupervised domain adaptive 3D object detection from point clouds. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), p. 6371–6380. External Links: Document Cited by: §I-A. [10] X. Peng, X. Zhu, and Y. Ma (2023) CL3D: unsupervised domain adaptation for cross-LiDAR 3D detection. In Proc. AAAI Conf. Artif. Intell., Vol. 37, p. 2047–2055. External Links: Document Cited by: §I-A. [11] Q. Hu, D. Liu, and W. Hu (2023) Density-insensitive unsupervised domain adaption on 3D object detection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 17556–17566. External Links: Document Cited by: §I-A, TABLE I. [12] Z. Chen, Y. Luo, Z. Wang, M. Baktashmotlagh, and Z. Huang (2023) Revisiting domain-adaptive 3D object detection by reliable, diverse and class-balanced pseudo-labeling. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), p. 3691–3703. External Links: Document Cited by: §I-A, TABLE I, TABLE I, TABLE I, TABLE I, TABLE I, TABLE IV. [13] Z. Zhang, M. Chen, S. Xiao, L. Peng, H. Li, B. Lin, P. Li, W. Wang, B. Wu, and D. Cai (2024) Pseudo label refinery for unsupervised domain adaptation on cross-dataset 3D object detection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 15291–15300. Cited by: §I-A, TABLE I. [14] X. Lu and H. Radha (2024) DALI: domain adaptive LiDAR object detection via distribution-level and instance-level pseudo label denoising. IEEE Trans. Robot. 40, p. 3866–3878. External Links: Document Cited by: §I-A, TABLE I. [15] S. Shin, Y. He, X. Hou, S. Hodgson, A. Markham, and N. Trigoni (2025) DiffRefine: diffusion-based proposal specific point cloud densification for cross-domain object detection. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), p. 4888–4897. Cited by: §I-A. [16] D. Hegde and V. M. Patel (2024) Attentive prototypes for source-free unsupervised domain adaptive 3D object detection. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), p. 3066–3076. Cited by: §I-A. [17] G. Chang, W. Roh, S. Jang, D. Lee, D. Ji, G. Oh, J. Park, J. Kim, and S. Kim (2024) CMDA: cross-modal and domain adversarial adaptation for LiDAR-based 3D object detection. In Proc. AAAI Conf. Artif. Intell., Vol. 38, p. 972–980. External Links: Document Cited by: §I-A, TABLE I. [18] S. Zhao, P. Liang, and Z. Yang (2026) Multi-modal assistance for unsupervised domain adaptation on point cloud 3D object detection. In Proc. AAAI Conf. Artif. Intell., Vol. 40, p. 13226–13234. External Links: Document Cited by: §I-A. [19] C. R. Qi, H. Su, K. Mo, and L. J. Guibas (2017) PointNet: deep learning on point sets for 3D classification and segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 652–660. Cited by: §I-B. [20] C. R. Qi, L. Yi, H. Su, and L. J. Guibas (2017) PointNet++: deep hierarchical feature learning on point sets in a metric space. In Adv. Neural Inf. Process. Syst., Vol. 30, p. 5099–5108. Cited by: §I-B. [21] R. Zhang, L. Wang, Y. Wang, P. Gao, H. Li, and J. Shi (2023) Starting from non-parametric networks for 3D point cloud analysis. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 5344–5353. Cited by: §I-B, Fig. 3, §I-B. [22] K. Chaney, F. Cladera, Z. Wang, A. Bisulco, M. A. Hsieh, C. Korpela, V. Kumar, C. J. Taylor, and K. Daniilidis (2023) M3ED: multi-robot, multi-sensor, multi-environment event dataset. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), p. 4016–4023. External Links: Document Cited by: §I-B. [23] M. K. Wozniak, M. Hansson, M. Thiel, and P. Jensfelt (2024) UADA3D: unsupervised adversarial domain adaptation for 3D object detection with sparse LiDAR and large domain gaps. IEEE Robot. Autom. Lett. 9 (12), p. 11210–11217. External Links: Document Cited by: §I-B. [24] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala (2019) PyTorch: an imperative style, high-performance deep learning library. In Adv. Neural Inf. Process. Syst., Vol. 32, p. 8024–8035. Cited by: §IV-A. [25] OpenPCDet Development Team (2020) OpenPCDet: an open-source toolbox for 3D object detection from point clouds. Note: [Online]. Available: https://github.com/open-mmlab/OpenPCDetAccessed: Aug. 25, 2026 Cited by: §IV-A. [26] S. Shi, C. Guo, L. Jiang, Z. Wang, J. Shi, X. Wang, and H. Li (2020) PV-RCNN: point-voxel feature set abstraction for 3D object detection. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), p. 10529–10538. Cited by: §IV-A. [27] J. Deng, S. Shi, P. Li, W. Zhou, Y. Zhang, and H. Li (2021) Voxel R-CNN: towards high performance voxel-based 3D object detection. In Proc. AAAI Conf. Artif. Intell., Vol. 35, p. 1201–1209. External Links: Document Cited by: §IV-A. [28] D. P. Kingma and J. Ba (2015) Adam: a method for stochastic optimization. In Proc. Int. Conf. Learn. Represent. (ICLR), Note: arXiv:1412.6980 Cited by: §IV-A. [29] L. N. Smith and N. Topin (2019) Super-convergence: very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, T. Pham (Ed.), Proc. SPIE, Vol. 11006. Note: Art. no. 1100612, doi: 10.1117/12.2520589 Cited by: §IV-A.