acceptodds
Withdrawn from ICLR 2027

GM3DMOT: Online 3D Graph-Mamba Multi-Object Tracking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Online LiDAR 3D multi-object tracking requires reliable association over sparse, unordered, and dynamically changing detections. Existing learning-based geometry-aware methods enrich association representations through sparse graph propagation or dense inter-object interactions. However, sparse graphs capture limited context, while dense interactions may introduce interference from weakly related objects. We propose GM3DMOT, an online LiDAR 3D MOT framework for identity-aware local and frame-level context modeling. Its Geometry-Conditioned Object Graph-Mamba Encoder comprises two complementary branches. The Same-Class Local Interaction Branch uses a same-class BEV graph with edge-gated message passing to model local ambiguity among nearby identity competitors. The Score-Degree Mamba Scan Branch combines detection confidence with same-class graph in-degree to serialize unordered detections for broader frame-level propagation. Scan-path geometric cues encode BEV distance and class consistency between consecutive detections in the serialized sequence. Before association decoding, a lightweight motion-conditioned relation refinement introduces short-term detection–track consistency. On nuScenes, GM3DMOT achieves 73.9% AMOTA with 138 identity switches on validation and 70.3% AMOTA with 198 identity switches on test.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Multi-Object Tracking, LiDAR, State Space Models, Graph Neural Networks, Data Association

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Qb9xOpenReview ↗
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)

Reviewer_XJMPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GM3DMOT addresses the challenge of reliable identity association in online LiDAR-based 3D multi-object tracking, where effectively modeling local object interactions and temporal dependencies is essential. To this end, it combines localized graph interactions with frame-level Mamba scanning, using detection confidence and node in-degree to determine the scanning order and integrating scan-path geometry with short-term motion cues. Experiments on nuScenes demonstrate competitive performance, achieving 0.739 AMOTA with 138 identity switches on the validation set and 0.703 AMOTA with 198 identity switches on the test set. Ablation studies further investigate the contributions of individual components and scanning strategies.

Critical strengths
  • The design is motivated by specific challenges in 3D MOT. To resolve identity ambiguity among nearby same-class objects while capturing broader contextual and temporal dependencies, the method combines same-class local graph interactions with frame-level Mamba scanning, incorporating scan-path geometry and short-term motion cues. Controlled experiments support the complementary roles of these components in reducing identity switches.
  • The experimental analysis is relatively comprehensive. Beyond benchmark comparisons, the paper provides three independent runs, component and ordering ablations, density-stratified analysis, transition-coverage normalization, and cross-detector evaluation. These analyses help identify the sources of improvement and reduce reliance on single-run results.
Critical weaknesses
  • The novelty lies primarily in component integration and task-specific adaptation, while the additional benefit of the core ordering mechanism is modest. The method builds on existing graph convolution, Mamba, and association decoding components. Compared with confidence-only ordering, score-degree ordering reduces average IDS only from 142 to 138, providing limited evidence for the distinctive advantage and general applicability of in-degree as an identity-association cue.
  • Performance gains involve trade-offs, and evidence of generalization remains limited. The Mamba variant has higher measured latency than the dense self-attention variant. Compared with the published GRAE-3DMOT results, some metrics, including MOTA and FRAG, deteriorate. Furthermore, evaluation is restricted to nuScenes, and cross-detector testing does not substitute for cross-dataset validation.
Questions for authors
  • Although Section 3.2 motivates the use of in-degree through local topological prominence, the paper does not directly analyze its relationship with association ambiguity after accounting for detection confidence and local density. This analysis is important for understanding the rationale and added value of Score-Degree ordering. Please clarify this connection through an analysis of association errors across in-degree levels or comparisons with alternative structural cues, such as neighborhood counts and average neighbor distance.
  • Compared with confidence-only ordering, Score-Degree ordering improves AMOTA from 0.737 to 0.739 and reduces IDS from 142 to 138, while MOTA remains unchanged at the reported precision. These gains are modest, and their distribution across different scenarios remains unclear. Please provide a breakdown by class, density, or occlusion level, together with scene-level uncertainty estimates, in Section 4.4 or Appendix F to clarify the sources and robustness of the improvements.
  • The evaluation is limited to nuScenes, leaving cross-dataset generalization insufficiently validated.
Reviewer_1uuROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GM3DMOT, an online LiDAR 3D multi-object tracker that combines same-class local graph interactions with Mamba-based frame-level context modeling. Detections are serialized according to confidence and graph in-degree, with additional scan-path geometric cues. The method builds on GRAE-3DMOT's association framework and introduces motion-conditioned relation refinement. Experiments on nuScenes achieve 73.9% AMOTA and 138 IDS on validation.

Critical strengths
  1. Competitive tracking performance. The method achieves strong results on nuScenes, particularly in reducing identity switches. Under controlled comparisons with GRAE-3DMOT, IDS decreases from 158 to 138 while AMOTA improves from 73.4% to 73.9%.
  2. Thorough ablations. The paper evaluates individual architectural components, scan-ordering strategies, and alternative context operators, including self-attention and GRU. The additional identity-switch and cross-detector analyses are appreciated.
Critical weaknesses
  1. Limited methodological novelty and unclear motivation for the architecture. The method largely combines existing graph-based interactions, Mamba, and motion-conditioned association features within the GRAE-3DMOT framework. Notably, Mamba operates only on current-frame detections, without modeling track histories or detection-track interactions. This asymmetry is insufficiently motivated given that temporal identity association is the central task. The improvements from the proposed score-degree ordering over confidence-only ordering are also relatively small.
  2. Limited experimental comparisons and modest overall gains. Several relevant recent approaches, including MCTrack and OptiPMB, are discussed but omitted from quantitative comparisons. The primary experiments are limited to CenterPoint detections on nuScenes. Although inference-only transfer to alternative detectors is evaluated, there is no independent training and evaluation with other detection backbones. Furthermore, test AMOTA improves by only 0.1 percentage points over GRAE-3DMOT, while fragmentation increases. Despite its efficiency motivation, Mamba is also slower than the evaluated dense self-attention alternative (18.01 vs. 15.63 ms/frame).
Questions for authors
  1. What motivates applying Mamba exclusively to current detections rather than track representations or joint detection-track interactions? Is there evidence that this asymmetric design is preferable?
  2. Can the authors provide comparisons with additional recent 3D MOT approaches, particularly MCTrack and OptiPMB, ideally using identical detections?
  3. Does Mamba offer any practical scalability benefit over self-attention as the number of detections increases? The current experiments show higher latency.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.