Research map / Graph learning / Knowledge and temporal graphs
Multi-Scene Generalized Trajectory Global Graph Solver with Composite Nodes for Multiple Object Tracking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoNo-Link, a graph-based global multi-object tracker that builds a partially connected graph over composite detection and trajectory nodes (chosen by NodeNet top-k) and refines it with a GNN link block, reducing memory versus fully connected graphs.
Problem
Graph-based global multi-object tracking uses fully connected graphs whose storage grows quadratically with clip length, and one-forward frame-to-frame associations miss long-range identity cues and lack end-to-end mismatch correction.
Approach
Replace fully connected frame-by-frame graphs with a partially connected graph whose nodes are composite (detection and trajectory) nodes, with edges selected by a NodeNet top-k query-based matcher, and a complementary edge learning strategy that re-learns trajectory-level edges so misclassified links can be recovered.
Key results
- On MOT17 private test, achieves 83.7 IDF1, 67.1 HOTA, 82.7 MOTA, outperforming SUSHI (83.1 IDF1).
- On MOT20 private test, 81.8 IDF1 and 77.5 MOTA, improving over ByteTrack by 6.4 IDF1 and 2.4 MOTA.
- On the MOT17 validation ablation, the partial graph G^part uses 143,603 average edges versus 256,095 for G^local_fully, with 3,523 MiB VRAM versus 3,689 MiB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoNo-Link instead of SUSHI | MOT17 tracking · MOT17 test, private detections | IDF1: 83.1 -> 83.7 | Helped |
| Use CoNo-Link instead of SUSHI | MOT20 tracking · MOT20 test, private detections | IDS: 706 -> 956 | Hurt |
| Use CoNo-Link instead of SUSHI | DanceTrack tracking · DanceTrack test | HOTA: 63.3 -> 63.8 | Helped |
| Use partially connected composite-node graph instead of hierarchical neighboring-frame fully connected graph (ablation) | MOT17 graph solver · MOT17 validation, GT detections | IDF1: 76.2 -> 76.3 | Helped |
| Use partially connected composite-node graph instead of fully connected graph (ablation) | GNN tracker · MOT17 validation | IDF1: 57.8 -> 58.1 | Helped |
| Add LS_1, mapping valid Traj-Traj edges to Det-Det edges (ablation) | Gpart + GNN · MOT17 validation | MOTA: 60.1 -> 59.9 | Hurt |
| Add LS_1, mapping valid Traj-Traj edges to Det-Det edges (ablation) | Gpart + GNN · MOT17 validation | IDF1: 58.1 -> 64.4 | Helped |
| Add LS_2, mapping Traj-Traj links and scores to Det-Det links (ablation) | Gpart + GNN · MOT17 validation | IDF1: 58.1 -> 61.3 | Helped |
| Add LS_3, mapping Traj-Traj links and re-extracting scores with GNN (ablation) | Gpart + GNN · MOT17 validation | IDF1: 58.1 -> 65.2 | Helped |
| Add LS_4, directly fusing trajectory nodes using Traj-Traj links (ablation) | Gpart + GNN · MOT17 validation | IDF1: 58.1 -> 69.9 | Helped |
Surprising findings
- On MOT20 the method slightly underperforms SUSHI in identity switches (956 vs 706 IDS).
- Trajectory-level learning strategy variants can slightly decrease MOTA (LS_1) while improving IDF1 and IDS.
Limitations
- Performance degrades when tracking targets after long occlusions because no correct association exists within the Traj-Traj links.
- Slightly weaker than SUSHI on MOT20 identity switches in dense scenes.
- Tracker running on graph construction is implemented in Python, slowing the overall system.
Experimental setup
Datasets: MOT17 MOT20 DanceTrack
Benchmarks: MOT17 MOT20 DanceTrack
Metrics: IDF1 MOTA HOTA IDS MT AssA DetA FLOPs FPS VRAM
Baselines: SUSHI ByteTrack MOTR GTR MPNTrack Lif_T MeMOT TrackFormer FairMOT GRTU Unicorn CorrTracker
Builds on: message-passing graph network tracker (MPNTrack) network flow MOT formulation SUSHI DETR-style query decoding focal loss
Closest papers by meaning
- Unifying Short and Long-Term Tracking With Graph Hierarchies (CVPR 2023 · similarity 0.89)
- Learnable Graph Matching: Incorporating Graph Partitioning With Deep Feature Learning for Multiple Object Tracking (CVPR 2021 · similarity 0.89)
- Sampling-Resilient Multi-Object Tracking (AAAI 2024 · similarity 0.87)
- Focusing on Tracks for Online Multi-Object Tracking (CVPR 2025 · similarity 0.87)
- DeconfuseTrack: Dealing with Confusion for Multi-Object Tracking (CVPR 2024 · similarity 0.86)
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.