- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
CuraScene selects training scenarios for motion forecasting using inferred interaction graphs. It learns sparse directed graphs from agent trajectories, extracts ego-centric subgraphs, and selects scenarios according to their structural novelty. At a reported 50% data budget, the selected subsets improve SeNeVa and DeMo over random subsampling on Waymo and Argoverse 2. The paper also evaluates agreement between the recovered agent sets and a relevance proxy derived from Waymo annotations.
Critical strengths
The problem is practically relevant. Using interaction structure to identify redundant driving scenarios is an interesting idea, and the selection method can be used without changing the downstream forecasting model.
Critical weaknesses
The experiments do not establish the main claimed contribution. Random sampling is the only selection baseline, and all downstream results use a single 50% budget. There is no comparison with embedding-based diversity selection or a simple interaction-density heuristic. More importantly, Table 2 is not a downstream selection ablation: it measures agreement with a set of relevant agents, rather than forecasting accuracy after training on subsets selected by the ablated methods. It also does not evaluate edge direction. Consequently, the results do not separate the value of learned directed topology from ordinary deduplication or preferential selection of more interactive scenes. Since the paper explicitly places its contribution in the selection signal rather than a new graph-learning algorithm, this missing evidence weighs heavily in my assessment.
The matched-budget experiment cannot be fully reconstructed from the reported protocol. Section 5.1 states that both methods select exactly half of the original scenarios. Algorithm 1, however, first removes duplicate graphs and then calibrates approximately 50% retention, without explaining how this produces the required number of original scenarios. The Argoverse 2 split also needs clarification: Appendix B describes the 224,896 scenarios as the combined training and validation splits, while the main experiment uses this number to describe the pool without identifying the held-out evaluation set. Actual split and subset counts are needed to verify the comparison. The paper acknowledges several limitations, but these details concern the validity of its main result.
The comparison with existing work is not convincing enough to establish the contribution. Most of the related-work discussion focuses on work published in 2024 or earlier, and the evaluation uses SeNeVa and DeMo, both from 2024. More importantly, these are downstream forecasting models; the only baseline for the proposed selection strategy is uniform random sampling. Beating random sampling does not establish an advantage over existing diversity-based or coreset selection methods. Given that the graph-learning and selection components largely build on established techniques, this missing comparison makes it difficult to assess the novelty and practical value of their combination. The authors should engage more closely with recent relevant work, including applicable 2026 methods available before submission, and justify their choice of baselines. As presented, the evidence for an advance over competitive selection approaches is insufficient.
I lean toward rejection. The gains over random sampling are promising, but they do not yet establish the value of the proposed interaction-based selection criterion. The missing downstream controls and the unclear budget and split accounting concern the paper’s main evidence, so I would need these issues resolved before supporting acceptance.
Questions for authors
Please provide the Argoverse 2 split sizes used for graph training, scenario selection, downstream training, and evaluation. After graph deduplication, how many scenarios remain, and how is the exact 50% budget enforced for both methods?
What node attributes and matching constraints are used in canonicalization and graph edit distance? In Appendix E(b), exchanging two otherwise indistinguishable agents A and B leaves a rooted graph with one incoming and one outgoing edge isomorphic to the original. Under a topology-only distance, how are the two examples distinguished?
Are any downstream results available for selection using the heuristic prior alone, undirected graphs, or generic feature diversity? Please also report the number of repeated runs and the variability behind Table 1.