PCSTracker: Long-Term Scene Flow Estimation for Point Cloud Sequences

Min Lin; Gangwei Xu; Xianqi Wang; Yuyi Peng; Xin Yang

cs.CVMar 20, 2026

PCSTracker: Long-Term Scene Flow Estimation for Point Cloud Sequences

Authors: Min Lin, Gangwei Xu, Xianqi Wang, Yuyi Peng, Xin Yang

Organizations: Huazhong University of Science and Technology · Optics Valley Laboratory

Abstract

Point cloud scene flow estimation is fundamental to long-term and fine-grained 3D motion analysis. However, existing methods are typically limited to pairwise settings and struggle to maintain temporal consistency over long sequences as geometry evolves, occlusions emerge, and errors accumulate. In this work, we propose PCSTracker, the first end-to-end framework specifically designed for consistent scene flow estimation in point cloud sequences. Specifically, we introduce an iterative geometry motion joint optimization module (IGMO) that explicitly models the temporal evolution of point features to alleviate correspondence inconsistencies caused by dynamic geometric changes. In addition, a spatio-temporal point trajectory update module (STTU) is proposed to leverage broad temporal context to infer plausible positions for occluded points, ensuring coherent motion estimation. To further handle long sequences, we employ an overlapping sliding-window inference strategy that alternates cross-window propagation and in-window refinement, effectively suppressing error accumulation and maintaining stable long-term motion consistency. Extensive experiments on the synthetic PointOdyssey3D and real-world ADT3D datasets show that PCSTracker achieves the best accuracy in long-term scene flow estimation and maintains real-time performance at 32.5 FPS, while demonstrating superior 3D motion understanding compared to RGB-D-based approaches.

Figures & tables

Figure 1: Overview of PCSTracker pipeline. Given a point cloud sequence, we first divide it into partially overlapping sliding windows. Within each window, query points are initialized via a KNN-based interpolation to form initial trajectories. Local geometric similarities are then computed through point–voxel correlations and processed by a spatio-temporal point Transformer to capture long-range spatial–temporal dependencies. Finally, trajectories are iteratively refined through $K$ optimization steps.

Figure 2: Auxiliary Points. We visualize the auxiliary points selected by different sampling strategies, and Tab. 3 reports their impact on inference performance. Green points denote the selected auxiliary points; (a) shows point clouds colorized with RGB information, and (b) highlights one of the query points in red .

Method	Input	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{0.10}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.20}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.40}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.80}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{avg}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
SpatialTracker [ 52 ]	RGB-D	0.924	21.12	33.03	48.06	66.81	42.25	49.54	0.896
DELTA [ 25 ]	RGB-D	0.780	40.43	55.00	65.89	74.67	58.99	67.16	0.762
SceneTracker [ 40 ]	RGB-D	0.204	61.95	75.93	86.18	93.85	79.48	87.98	0.200
SF-baseline	Point	0.330	28.92	51.77	74.88	91.03	61.65	77.78	0.319
PCSTracker(Ours)	Point	0.133	70.19	84.48	93.37	97.44	86.37	93.65	0.129

Table 1: Comparison results on the PointOdyssey3D. We re-evaluate SpatialTracker [ 52 ] , SceneTracker [ 40 ] , and DELTA [ 25 ] on the corresponding RGB-D sequences of the PointOdyssey3D dataset using their publicly available code and pretrained checkpoints. In addition, we retrain the SF-baseline on PointOdyssey3D for a fair comparison. Our method achieves a reduction of 34.8% and 59.7% in

\text{EPE}_{3D}

compared with SceneTracker and SF-baseline, respectively, demonstrating a substantial improvement in the estimation performance.

Figure 3: Qualitative Results on PointOdyssey3D. The first column shows the input point cloud sequence (colored by RGB for visualization) and the corresponding query points. Columns 2–7 compare ground-truth trajectories with predictions from different methods. RGB-D-based methods [ 52 , 40 , 25 ] , constrained by their 2D appearance-driven frameworks, fail to recover reliable 3D trajectories, reflecting limited understanding of scene geometry and 3D motion. SF-baseline, which lacks a dedicated design for long sequences, suffers from significant errors. In contrast, our method produces more correct 3D motion understanding and accurate 3D trajectories.

Method	Input	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{0.10}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.20}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.40}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{0.80}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\delta_{3D}^{avg}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
SpatialTracker [ 52 ]	RGB-D	0.916	23.71	38.04	58.16	80.29	50.05	60.31	0.505
DELTA [ 25 ]	RGB-D	1.509	45.53	70.89	81.67	89.59	71.92	83.41	1.278
SceneTracker [ 40 ]	RGB-D	0.601	43.15	62.44	79.46	90.91	68.99	80.40	0.271
SF-baseline	Point	0.945	12.32	29.52	49.13	70.99	40.49	51.61	0.702
PCSTracker(Ours)	Point	0.372	46.73	71.74	85.49	93.81	74.44	87.74	0.226

Table 2: Generalization results on ADT3D dataset. We evaluate our method, SF-baseline, and RGB-D-based approaches on the ADT3D dataset and its corresponding RGB-D sequences. Our method reduces the

\text{EPE}_{3D}

by 38.1% and 60.6% compared with SceneTracker [ 40 ] and SF-baseline, respectively, demonstrating its superior generalization ability and robustness in realistic dynamic scenes.

Figure 4: Visualization of generalization results on ADT3D. Each row shows the point clouds and the motion trajectories of the query points at different timestamps within a given scene (each point cloud is colored with its RGB information for better visualization). More results are provided in the appendix.

Query	Support	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{\text{avg}}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
One	-	0.852	47.49	57.31	0.834
One	Loc.KNN	0.148	83.00	94.35	0.146
One	Glob.FPS	0.134	87.64	95.05	0.136
One	Glob.Random	0.164	85.64	93.40	0.164
One	Loc.KNN + Glob.Random	0.152	86.42	95.16	0.151
One	Loc.KNN + Glob.FPS	0.119	87.64	95.74	0.115

Table 3: Ablation of inference mode. We evaluate different inference modes on the PointOdyssey3D dataset to analyze how various auxiliary point selection strategies affect performance.

	Experiment	Variation	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{avg}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$
(a)	Geometry–Motion joint Optimization	w/o	0.202	75.85	90.61
	Geometry–Motion joint Optimization	w/	0.133	86.37	93.65
(b)	Window Size	2	0.206	78.33	90.68
		8	0.166	83.06	92.67
		16	0.133	86.37	93.65
(c)	Spatial-Temporal Transformer Block’s Number	1 $\times$ 2	0.185	82.93	92.92

Table 4: Results of ablation studies on the PointOdyssey3D.

Dataset	Method	Number of Frames
Dataset	Method	2	8	24	40
POD3D	SF-baseline	0.0320	0.165	0.385	0.543
POD3D	Ours	0.0237	0.0719	0.157	0.205
ADT3D	SF-baseline	0.0363	0.1746	0.3999	0.5906
ADT3D	Ours	0.0199	0.0605	0.1261	0.1854

Table 5: Comparison temporal drift across different frames. POD3D denotes the PointOdyssey3D dataset.

Method	Parameters	Runtime	FPS
SpatialTracker	34.0M	4.73s	8.46
SceneTracker	24.2M	1.38s	29.0
Ours	3.48M	1.23s	32.5

Table 6: Parameters and runtime comparison. Runtime comparison under the 40 frames, 1024 query points setting.

Dataset	Num. Iteration	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{\text{avg}}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
	1	0.171	81.16	91.75	0.167
	2	0.136	86.09	94.41	0.130
	4	0.133	86.37	93.65	0.129
POD3D	6	0.124	87.70	95.05	0.118
	8	0.126	87.52	95.02	0.120
	12	0.132	86.64	94.93	0.125

Table 7: Comparison results on PointOdyssey3D (POD3D) and ADT3D datasets with different iteration numbers.

Dataset	Num. Query	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{\text{avg}}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
	1	0.668	56.04	67.66	0.619
	32	0.154	83.13	93.72	0.148
	64	0.143	85.24	93.99	0.137
POD3D	256	0.142	84.84	94.01	0.138
	512	0.139	85.16	94.34	0.135
	1024	0.133	86.37	93.65	0.129

Table 8: Comparison results on PointOdyssey3D (POD3D) and ADT3D datasets with different numbers of query points.

Correlation Volume	$\text{EPE}_{3D}{\color[rgb]{1,0,0}\downarrow}$	$\delta_{3D}^{\text{avg}}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{Survival}_{3D}^{0.50}{\color[rgb]{0.2,0.8,0.2}\uparrow}$	$\text{MAE}_{3D}{\color[rgb]{1,0,0}\downarrow}$
Point-only	0.501	51.07	65.16	0.510
Voxel-only	0.150	84.95	93.22	0.145
Point-Voxel	0.133	86.37	93.65	0.129

Table 9: Impact of point-voxel correlation

Dataset	Method	Number of Occlusion Frames
Dataset	Method	0	2	4	$>$ 4
POD3D	SF-baseline	0.1950	0.2421	0.2714	0.3511
POD3D	Ours	0.0754	0.1034	0.1200	0.1736
ADT3D	SF-baseline	0.3404	0.4188	0.4690	0.5233
ADT3D	Ours	0.0961	0.1228	0.1303	0.1464

Table 10: Comparison performance under different occlusion levels. POD3D denotes the PointOdyssey3D dataset.

Figure 5: Appendix Results on ADT3D. Each row shows the point clouds and the motion trajectories of the query points at different timestamps within a given scene (each point cloud is colored with its RGB information for better visualization).

Figure 6: Spatial distribution gap across datasets. red points denote the training dataset (POD3D), while black points represent the datasets used for generalization evaluation.

Metric	DriveTrack3D	ADT3D
$\text{EPE}_{3D}(m)$	9.5	0.372

Table 11: Generalization Results on Different Datasets.

Figure 7: Appendix Results on ADT3D. Each row shows the point clouds and the motion trajectories of the query points at different timestamps within a given scene (each point cloud is colored with its RGB information for better visualization).

Figure 8: Appendix Results on ADT3D. Each row shows the point clouds and the motion trajectories of the query points at different timestamps within a given scene (each point cloud is colored with its RGB information for better visualization).

Figure 9: Appendix Results on PointOdyssey3D. The first column shows the input point cloud sequence (colored by RGB for visualization) and the corresponding query points. Columns 2–7 compare ground-truth trajectories with predictions from different methods.

Scene Flow Estimation

See all 12 papers

FocusTrack: One-Stage Focus-and-Suppress Framework for 3D Point Cloud Object Tracking
Feb 27, 2026Sifan Zhou, Jiahao Nie, Ziyu Zhao +2Visual Object Tracking Point Cloud Learning
School of Automation Southeast University, Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education Nanjing, China · School of Information Technology and Artificial Intelligence, Zhejiang University of Finance and Economics Hangzhou, China · Big Data Institute Central South University Changsha, China
PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences
Mar 22, 2026Lanbo Xu, Liang Guo, Caigui Jiang +1Streaming 3D Reconstruction Monocular 3D Reconstruction
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · University of East Anglia, Norwich, NR47TJ, United Kingdom
MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting
Oct 7, 2026Jiuming Liu, Jianing Li, Mengmeng Liu +3Scene Flow Estimation Spatiotemporal Forecasting
University of Cambridge · Shanghai Jiao Tong University · University of Twente +1

Feb 27, 2026J/K move · Enter open · S save

FocusTrack: One-Stage Focus-and-Suppress Framework for 3D Point Cloud Object Tracking

Sifan Zhou, Jiahao Nie, Ziyu Zhao, Yichao Cao+1

School of Automation Southeast University, Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education Nanjing, China · School of Information Technology and Artificial Intelligence, Zhejiang University of Finance and Economics Hangzhou, China · Big Data Institute Central South University Changsha, China

Visual Object Tracking Point Cloud Learning Attention Mechanisms

Open paper AskPDF arXiv

In 3D point cloud object tracking, the motion-centric methods have emerged as a promising avenue due to its superior performance in modeling inter-frame motion. However, existing two-stage motion-based approaches suffer from fundamental limitations: (1) error accumulation due to decoupled optimization caused by explicit foreground segmentation prior to motion estimation, and (2) computational bottlenecks from sequential processing. To address these challenges, we propose FocusTrack, a novel one-stage paradigms tracking framework that unifies motion-semantics co-modeling through two core innovations: Inter-frame Motion Modeling (IMM) and Focus-and-Suppress Attention. The IMM module employs a temp-oral-difference siamese encoder to capture global motion patterns between adjacent frames. The Focus-and-Suppress attention that enhance the foreground semantics via motion-salient feature gating and suppress the background noise based on the temporal-aware motion context from IMM without explicit segmentation. Based on above two designs, FocusTrack enables end-to-end training with compact one-stage pipeline. Extensive experiments on prominent 3D tracking benchmarks, such as KITTI, nuScenes, and Waymo, demonstrate that the FocusTrack achieves new SOTA performance while running at a high speed with 105 FPS.

Abstract

Scene Flow Estimation

Explore similar work

FocusTrack: One-Stage Focus-and-Suppress Framework for 3D Point Cloud Object Tracking

PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences

MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting