- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents GM3DMOT, an online Multi-Object Tracking (MOT) method in LiDAR. The model inputs 3D detections, embeds them with an MLP and refines them through: A message-passing Graph Neural Network (GNN) that connects each detection with its k-Nearest Neighbors from the same class. A Mamba block, for which the scanning order is determined by the confidence and the number of incoming edges of the detections. The refined detection features, the existing track features, as well as a geometry/motion-based similarity matrix between the detections and the tracks are fed to the same decoder and matching procedure as GRAE-3DMOT. In a controlled comparison against GRAE-3DMOT, they obtain better AMOTA 0.734 → 0.739 and they reduce Identity Switches (IDS) from 158 to 138.
Critical strengths
S1. The controlled comparison with GRAE-3DMOT. In Table 1, the authors present a controlled comparison between their method and their reproduction of GRAE-3DMOT (using the same detections and training setup). This comparison shows that on the nuScenes val set, their method improves AMOTA from 0.734 to 0.739, and reduces IDS from 158 to 138.
S2. The large number of ablations. Throughout the main paper and the supplementary, the authors present many experiments and ablations: Mamba vs Dense Self-Attention vs GRU, Mamba scan ordering, generalization to other detectors, neighborhood size k for the graph edges construction, performance stability across multiple runs, etc… Some of these ablations could be valuable for the community.
Critical weaknesses
W1. Limited gains. Despite the architectural complexity brought by the GNN and the Mamba blocks, GM3DMOT only brings limited performance improvements over the GRAE-3DMOT baseline. In the controlled comparison on the nuScenes val set, their method AMOTA from 0.734 to 0.739, and reduces IDS from 158 to 138. The gap is even smaller when comparing to the published GRAE-3DMOT results.
W2. Limited novelty. Their baseline GRAE-3DMOT already refines detection features using geometric relations between objects, and GM3DMOT uses the same association decoder and matching procedure as GRAE-3DMOT. The specific contribution is thus the combination of same-class local message passing, Mamba scanning, and motion-conditioned refinement. These are coherent choices, but the paper would be stronger if it showed that this design was transferable to other pipelines and/or if it presented a substantial advantage over alternative feature refinement methods.
Questions for authors
Q1. If doable, I think training and reporting performance on the Waymo Open Dataset - like CenterPoint and DSC-Track do - would strengthen the paper.
Q2. The ablation of the Mamba scan ordering in Table 3 is interesting. However the evaluated baselines seem quite arbitrary, especially since the scan ordering is claimed as a contribution. I would like to see other commonly used scan ordering being tested, e.g. a Hilbert-index ordering, as in [1], and/or a simple scan ordering based on the coordinates of the centers of the detections [2] (either in BEV or in 3D).
Q3. I was surprised to see in Table 2 and Table 15 that the dense self-attention variant is faster but worse. The paper already explained that the speed difference is due to the short-sequence regime, but are there any insights as to why attention performs worse here?
[1] G. Zhang et al, “Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection” (NeurIPS’24) [2] Z. Liu et al, “LION: Linear Group RNN for 3D Object Detection in Point Clouds” (NeurIPS’24)