- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposed CooperTrace, which have a specific position in the intersection of cooperative, multi-modal, and end-to-end motion prediction. Extensive evaluations on several cooperative datasets show that CooperTrace's good performance. The paper also provides stress-test performance on cooperative gains under emulated sensor degradation, which is meaningful for real-world applications.
Critical strengths
- This paper provides valuable empirical analysis on the tracking-to-forecasting distribution gap. It proposes an explicit tracking history, and demonstrate the different effects of subcomponents. The findings provide transferable design insights for other motion forecasting systems, which may benefit even without adopting CooperTrace model.
- The evaluation is relatively comprehensive with several real-world and simulated datasets. And authors use more comprehensive metrics like coverage, EPA, etc., beyond naive ADE, especially the same-population analysis helps demonstrate model capability in different-difficulty cases.
Critical weaknesses
This paper positions end-to-end framework as one of the core contributions, but Table 6 shows that joint fine-tuning only improves minADE from 1.365 to 1.338 m. And the main OPV2V result in Table 2 appears to correspond to the staged LiDAR-camera configuration in Table 6, rather than the jointly fine-tuned model. Why is the jointly fine-tuned multi-modal model not used for the main results? Overall, although the authors demonstrate that joint backpropagation is feasible, but current experiments cannot sufficiently show that such end-to-end backpropagation is the key source of performance gains.
Questions for authors
- Following the mentioned critical weakness, can authors provide the end-to-end forecasting ablations (particularly the contribution of joint fine-tuning) on the other real-world datasets? Current experiment results may be not distinguishing enough to show the separate contributions.
- Given the large improvement from explicit tracking histories in Table 7, how well would a history-only forecaster perform? In particular, it would be useful to know how much additional benefit the track queries and BEV features provide once the motion history is available.
- How are the low, moderate, and high severity levels defined for the "Camera missing" and "LiDAR missing" corruptions? The implementation and corruption parameters for each level are not clearly specified.
2 comments
Thank you for the detailed response.
For W1 and Q1, the additional comparisons clarify the contribution of end-to-end joint training, although the improvements remain marginal.
For Q2, the new experiments convincingly demonstrate the importance of explicit history, and I appreciate the authors' honest reporting, although the results also somewhat weaken the contribution of track queries.
Overall, my assessment remains somewhat critical. Nevertheless, I still find the authors' exploration and transparent empirical findings meaningful, and I will maintain my original rating.
Thank you for the careful review, for recognizing our analysis of the tracking-to-forecasting gap, and for the focused questions on the end-to-end claim. We address each point below. Changes in the revised PDF are in blue.
1. The contribution of joint fine-tuning, and the model reported in Table 2 (W1, Q1)
Please see point 2 of the general response. For the joint fine-tuning ablation on real-world data, the revised jointly fine-tuned model on V2X-Seq improves over staged training in minADE@5s (1.036 → 1.019 m), coverage (0.780 → 0.803) and EPA (0.471 → 0.480). Table 2 now reports the jointly fine-tuned model.
In the revised PDF, Sec. 1 is reworded, Table 2 reports the jointly fine-tuned model, Table 8 keeps the cached-history row, and Appendix I reports the staged and jointly fine-tuned numbers.
2. A history-only forecaster, and what the track query and BEV features add (Q2)
Please see point 1 of the general response. In short, a history-only forecaster already performs well (minADE@5s of 1.295 m in the cooperative setting), BEV features contribute about 0.11 m, and the track query raises EPA from 0.422 to 0.450 once it is fused through a learned gate.
In the revised PDF, Table 9 holds the cooperative rows and appendix Table 22 the single-agent ones, Sec. 3.4 describes the gated fusion, and Sec. 4.5 explains the result.
3. Severity levels of the corruptions (Q3)
At low / moderate / high severity, "camera missing" blanks each camera view (the whole image set to zero) independently with probability 0.3 / 0.7 / 1.0, so at high severity no camera image remains, and "LiDAR missing" keeps a random 10% / 3% / 1% of the points. The new appendix Table 21 lists every corruption.
In the revised PDF, Appendix D lists every corruption's parameters per severity.
Summary. We revised joint fine-tuning so that the forecaster trains on live tracker outputs, repeated the ablation on real-world V2X-Seq and report the jointly fine-tuned model in Table 2, added the history-only forecaster and a gated query fusion, and specified the severity levels. We would be glad to answer any follow-up questions.