Research map / Detection and segmentation / Tracking
Sampling-Resilient Multi-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 SR-Track, a sampling-resilient multi-object tracker using an LSTM-based sparse-observation Kalman filter with Bayesian-NN noise estimation and a combined spatial matching metric, to keep tracking accurate under down-sampled frames.
Problem
Multi-object tracking degrades sharply when only a small subset of video frames is processed (high frame reduction ratios), since motion becomes non-linear and IoU-based association fails on sparse frames.
Approach
Replace the linear state transition of the Kalman filter with an LSTM that takes the observation and temporal gap as input, estimate process and observation noise with Bayesian neural networks, derive an optimal Kalman gain, and associate boxes with a combined overlap, center-distance and aspect-ratio metric.
Key results
- On MOT17 at frame reduction ratio 9, SR-Track reaches HOTA 63.4 versus ByteTrack's 58.9 at 29.4 FPS.
- With the same HOTA of 62 on MOT17, SR-Track needs 17.9s versus ByteTrack's 36.9s, and on DanceTrack the speedup is about 3x.
- On the DanceTrack test set without down-sampling, SR-Track reaches HOTA 59.1 versus OC-SORT's 55.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SR-Track instead of ByteTrack | MOT tracking · MOT17, RR = 9 | HOTA: 58.9 -> 63.4 | Helped |
| Use SR-Track instead of OC-SORT | MOT tracking · MOT20, RR = 9 | HOTA: 50.7 -> 55.8 | Helped |
| Use SR-Track instead of OC-SORT | MOT tracking · DanceTrack, RR = 9 | HOTA: 33.4 -> 39.1 | Helped |
| Use SR-Track instead of OC-SORT | MOT tracking · DanceTrack test | HOTA: 55.1 -> 59.1 | Helped |
| Add SOKF (ablation) | SR-Track · MOT17, RR = 9 | HOTA: 59.6 -> 63.6 | Helped |
| Add RDA (ablation) | SR-Track · MOT17, RR = 9 | HOTA: 61.4 -> 63.6 | Helped |
| Add LSTM-based position prediction (ablation) | SOKF · MOT17, RR = 9 | HOTA: 60.8 -> 63.6 | Helped |
| Add BNN-based noise estimation (ablation) | SOKF · MOT17, RR = 9 | HOTA: 62.0 -> 63.6 | Helped |
| Add optimal Kalman gain (ablation) | SOKF · MOT17, RR = 9 | HOTA: 61.0 -> 63.6 | Helped |
Surprising findings
- SCALE (aspect-ratio) cue becomes more useful than IoU and center distance as the frame reduction ratio increases.
Limitations
- Expensive trackers could not be re-trained on DanceTrack due to hardware limits, so comparisons there are missing.
- Evaluation uses the annotated halves of training videos rather than the unannotated official test sets.
- Relies on a shared YOLOX detector and does not evaluate appearance features.
Experimental setup
Datasets: MOT17 MOT20 DanceTrack
Benchmarks: MOT17 MOT20 DanceTrack
Metrics: HOTA MOTA IDF1 FPS Time@HOTA AssA DetA
Baselines: ByteTrack OC-SORT SimpleTrack TransTrack TrackFormer MOTR StrongSORT BoT-SORT FairMOT QDTrack
Builds on: ByteTrack SORT OC-SORT
Closest papers by meaning
- DiffusionTrack: Diffusion Model for Multi-Object Tracking (AAAI 2024 · similarity 0.89)
- UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation (AAAI 2024 · similarity 0.87)
- DiffMOT: A Real-time Diffusion-based Multiple Object Tracker with Non-linear Prediction (CVPR 2024 · similarity 0.87)
- Multi-Scene Generalized Trajectory Global Graph Solver with Composite Nodes for Multiple Object Tracking (AAAI 2024 · similarity 0.87)
- More Than Meets the Eye: Enhancing Multi-Object Tracking Even with Prolonged Occlusions (ICML 2025 · similarity 0.87)
- Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking (AAAI 2024 · similarity 0.87)
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.