- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies source‑only cross‑sensor 3D single‑object tracking (3D‑SOT) for LiDAR point cloud. The authors observe that sensor shift brings two coupled distribution shifts: observation reliability (sparse / missing target points) and transition scale (variable time interval between annotated frames).
They propose ReLiTrack, an observation‑transition factorized tracker. Key components include:
- continuous‑time mixture‑of‑experts motion prior with moment‑matching;
- masked point encoding plus permanent anchor‑recent memory for appearance correction;
- covariance‑intersection fusion handling correlated motion‑appearance errors;
- paired range‑density degradation training to improve generalization under point sparsity.
Evaluations are conducted over all six directed source‑only transfer pairs across KITTI, Waymo, nuScenes for Vehicle and Pedestrian. ReLiTrack outperforms prior trackers on average Success / Precision metrics. Ablations validate each main module.
Critical strengths
-
Well‑defined cross‑sensor source‑only transfer setting for LiDAR 3D‑SOT. The paper quantitatively characterizes two key shifts (observation sparsity & time‑variant motion step) across three major autonomous driving datasets, and evaluates full six‑direction transfer protocol, which is more rigorous than common single source‑target pair comparison. This benchmark setup is valuable for future domain‑generalized 3D tracking research.
-
Factorized architecture design separating continuous‑time motion transition and observation‑driven appearance correction. The continuous‑time motion mixture expert and covariance‑intersection fusion bring clear empirical gains under large sensor / cadence changes, and ablation confirms the large performance contribution of these two modules.
Critical weaknesses
-
Missing mandatory AI Use Statement against ICLR‑2027 policy. Visual inspection suggests Figure 1 and Figure 3 bear typical AI‑generated plotting artifacts. The submission neither contains an AI Use Statement in main text / appendix, nor fills the corresponding OpenReview form. ICLR requires authors to disclose substantial LLM / AI usage for writing, plotting or figure creation. Missing this mandatory disclosure is a policy violation and raises transparency risk. This is a major concern for my weak‑reject recommendation.
-
Limited real‑world failure analysis. Although numeric cross‑dataset transfer gains are observed, the paper lacks qualitative case analysis showing when and why ReLiTrack still fails under severe sensor shift. We cannot fully tell whether performance improvement comes from handling true distribution shift or dataset‑specific biases. Without visualized failure cases, readers cannot diagnose remaining bottlenecks for real‑world deployment.
Questions for authors
-
Did you use any generative‑AI / LLM tools for creating Figure 1 and Figure 3? Please strictly follow ICLR 2027 AI‑use policy: provide a complete AI Use Statement in appendix, and complete the AI‑usage questionnaire on OpenReview. Clearly list which parts (figure plotting, text drafting, formula typesetting etc.) involve AI assistance and describe human post‑editing procedures.
-
Could you supply several qualitative failure examples from cross‑sensor transfer evaluation? Please illustrate typical error modes (e.g. zero‑return target crop, large inter‑frame displacement) to explain where your method still breaks.
-
In ablation, fixed‑step constant‑velocity transition causes huge performance drop. Can you clarify whether baselines like MBPTrack / M2‑Track also adopt fixed‑step motion assumption? This is important to confirm whether your continuous‑time transition advantage is fairly compared.
-
The author omitted some recent studies when comparing tracking methods; it is recommended that the following be included: “Revisiting Siamese-Based 3D Single Object Tracking With a Versatile Transformer” (TPAMI'2025) and “TrackAny3D: Transferring Pretrained 3D Models for Category-unified 3D Point Cloud Tracking” (ICCV'2025).
4 comments
Responses to Reviewer XGwk (Part. 2)
Response to Question #3:
Thank you very much for this question. The fixed-step constant-velocity model in Table 2 is an internal transition ablation of ReLiTrack, not a reproduction of M²-Track or MBPTrack. It replaces the proposed four-expert transition with a single constant-velocity predictor using a fixed temporal step. Therefore, its substantial performance degradation should not be interpreted as a direct measurement of the advantage of elapsed-time conditioning over either baseline.
M²-Track estimates relative motion from point clouds in consecutive frames. Its motion transformation is predicted for a frame pair, rather than explicitly parameterized by the physical elapsed time as in Eq. (3) of ReLiTrack. The original M²-Track supplementary material also acknowledges sensitivity to changes in sampling frequency, particularly when training and evaluation use different frame rates.
MBPTrack follows a different formulation. It propagates historical point-cloud features and targetness information through memory and performs coarse-to-fine localization. It does not employ the same explicit constant-velocity transition as our ablation. Thus, characterizing MBPTrack as a fixed-step constant-velocity method would be inaccurate.
To isolate the contribution of elapsed-time conditioning more precisely, we have clarified the comparison already reported in Supplementary Table S5:
| Transition variant | W→N Success/Precision | W→K Success/Precision |
|---|---|---|
| Constant velocity, fixed step | 40.5 / 53.3 | 55.8 / 70.2 |
| Constant velocity, actual elapsed time | 41.4 / 54.9 | 57.8 / 71.3 |
| Four-mode transition, actual elapsed time | 50.2 / 61.3 | 71.2 / 84.3 |
The first two rows isolate the effect of actual elapsed-time conditioning within the constant-velocity family, yielding improvements of 0.9/1.6 points on W→N and 2.0/1.1 points on W→K. The additional improvement of the full transition also reflects the modeling of static, acceleration, and turning regimes, rather than elapsed-time conditioning alone.
Response to Question #4:
Thank you for suggesting these relevant studies. We have carefully reviewed both papers and expanded the related-work discussion to better position ReLiTrack within recent advances in 3D single-object tracking. Regarding “Revisiting Siamese-Based 3D Single Object Tracking With a Versatile Transformer” (TPAMI 2025), its official implementation and pretrained checkpoints are not publicly available to the best of our knowledge, preventing a faithful reproduction on our benchmark.
TrackAny3D focuses on transferring pretrained 3D representations to category-unified tracking, whereas ReLiTrack addresses source-only cross-sensor transfer, with particular emphasis on changes in observation reliability and temporal sampling. We have supplemented our evaluation with a reproduction of TrackAny3D using its publicly released implementation. The reproduced method has been evaluated under our CD3DSOT-v0.1 benchmark, covering the six directed transfers among KITTI, Waymo, and nuScenes for Vehicle/Car and Pedestrian. We have reproduced TrackAny3D using its publicly released implementation and evaluated it on our unified cross-dataset benchmark, we will include the experimental results of TrackAny3D in the final version of the manuscript. The additional results are presented in Table R6 below.
Table R6. Cross-dataset comparison with TrackAny3D on CD3DSOT-v0.1. Results are Success/Precision (%).
| Method | K→W | K→N | W→K | W→N | N→K | N→W | Mean |
|---|---|---|---|---|---|---|---|
| Vehicle/Car | |||||||
| TrackAny3D | 56.4/66.7 | 27.8/44.1 | 69.9/81.5 | 45.6/56.0 | 57.2/68.9 | 49.5/63.2 | 51.07/63.40 |
| ReLiTrack | 60.1/68.8 | 29.5/47.6 | 71.2/84.3 | 50.2/61.3 | 58.5/62.3 | 52.8/66.5 | 53.72/65.13 |
| Pedestrian | |||||||
| TrackAny3D | 38.2/55.4 | 25.1/40.6 | 65.4/90.0 | 37.9/53.1 | 44.3/65.8 | 46.8/61.7 | 42.95/61.10 |
| ReLiTrack | 40.8/59.1 | 24.5/39.2 | 64.5/90.6 | 40.7/57.5 | 46.1/66.9 | 49.9/64.3 | 44.42/62.93 |
Thank you once again for your insightful questions and suggestions.
Thank you for the reviewers’ careful feedback. Most of my concerns have been addressed, and I am willing to raise my score to 3.
Good luck!
Response to Reviewer XGwk
Dear Reviewer XGwk,
Thank you very much for your positive feedback and for reconsidering your assessment. We have supplemented the failure cases and visualization diagrams in the supplementary materials. In the final version of the supplementary materials, we will add more visualization results and improve the descriptions.
We sincerely appreciate your constructive comments, which have helped us improve the clarity and quality of our work.
Responses to Reviewer XGwk
Dear Reviewer XGwk
Thank you very much for your questions and suggestions, please find our responses as follows:
Response to Weakness #1 and Question #1:
Thank you very much for raising this important concern. We fully recognize the importance of transparency regarding AI-assisted research and sincerely apologize for the omission of the required AI Use Statement in the original submission. We have addressed this omission in the revised manuscript and clarified the specific roles of AI assistance in figure preparation and language polishing. We would particularly like to clarify that neither Figure 1 nor Figure 3 was directly synthesized by a generative image model. Both figures were rendered programmatically using Python and Matplotlib, with their scientific content determined by the data analysis or mathematical formulations.
For Figure 1, the six panels present actual statistical measurements from KITTI, Waymo, and nuScenes. The underlying data-processing procedures extract whole-frame and target-box point counts, range-conditioned observation statistics, inter-frame displacement, and object dimensions. The statistical analysis uses deterministic trajectory sampling and the original dataset annotations and point clouds. As documented in Supplementary Section S2, the acquisition statistics include track-balanced sampling, while the displacement analysis contains 15,062 KITTI, 25,854 Waymo, and 8,444 nuScenes transitions. The plots were constructed from these numerical measurements using Matplotlib. No statistical values or experimental observations were generated by AI.
For Figure 3, the plotted curves were constructed from the Gaussian mixture and moment-matching equations in our method. Specifically, the weighted Gaussian components and their mixture density were evaluated mathematically, and the corresponding single-Gaussian approximation was obtained from the mixture mean and variance. The resulting probability density curves were visualized using Matplotlib. This figure is explicitly an illustrative mathematical example, not an empirical result or an AI-synthesized image.
Regarding AI assistance, we independently developed the research methodology, mathematical derivations, experimental design, implementation, and original manuscript. A Nature-style AI-assisted plotting skill was used to facilitate figure preparation and improve visual presentation. AI tools were also used for manuscript language polishing. We have reviewed and manually refined the resulting figures and verified their numerical and mathematical content.
We have provided a complete AI Use Statement in our revised manuscript. We emphasize that the original omission concerned the reporting of AI assistance, rather than the authenticity of the data, mathematical formulations, or scientific results. We appreciate the reviewer’s attention to research transparency and hope that the corrected disclosure and detailed explanation of the figure-generation procedures have resolved this concern.
Response to Weakness #2 and Question #2:
Thank you very much for this valuable suggestion. As OpenReview does not support direct image uploads in the response, we have outlined the planned qualitative failure analysis below:
Zero-return targets: Analyze tracking drift when geometric evidence is unavailable.
Large inter-frame displacement: Examine failures when abrupt motion exceeds the predicted search region.
Dense distractors: Investigate incorrect target associations and subsequent memory drift.
We expect to complete these experiments and provide the corresponding visualizations within two days. Thank you for your understanding and patience.