---
license: mit
tags:
- multi-object-tracking
- re-identification
- tracking
---

# VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking of Highly Similar Objects

Weights and precomputed assets for **[VLA-ReID](https://github.com/holmescao/VLA-ReID)** (Video-Level Association re-ID): set-relative appearance re-scoring for multi-object tracking of highly similar objects, evaluated on the [BEE24 benchmark](https://holmescao.github.io/datasets/BEE24). Frame-Common Appearance Estimation (FCAE) + Common Appearance Suppression (CAS) plug into TrackTrack by replacing only its appearance term.

**BEE24 test (official 5-sequence split, shared YOLOX detections):**

| Method | HOTA | MOTA | IDF1 | AssA | IDs |
|---|---:|---:|---:|---:|---:|
| TrackTrack (baseline) | 48.05 | 65.69 | 64.12 | 45.91 | 606 |
| **VLA-ReID (ours)** | **49.22** | **67.02** | **64.93** | **46.64** | **437** |

See the [GitHub repository](https://github.com/holmescao/VLA-ReID) for installation and step-by-step reproduction instructions; the file-by-file manifest (destination paths + SHA-256) is in [WEIGHTS.md](https://github.com/holmescao/VLA-ReID/blob/main/WEIGHTS.md).

| File | Bytes | Role |
|---|---:|---|
| `yolox_x_bee24.pth.tar` | 792,835,731 | YOLOX-X bee detector |
| `bee24_test_0.80.pickle` / `bee24_test_0.95.pickle` | 6,162,932 / 8,818,700 | shared detections (dual NMS) |
| `sam3_bee24_finetuned.pt` | 3,371,888,621 | SAM3 segmentation checkpoint fine-tuned on bee foreground (base model by Meta AI; redistributed under the SAM license terms) |
| `bee_direction_resnet18_triclass.joblib` | 44,790,820 | bee orientation classifier |
| `reid_stage1_agw_s50.pth` | 331,487,317 | stage-1 AGW-ResNeSt50 encoder (baseline) |
| `reid_stage1_agw_s50_ep60.pth` | 331,487,317 | stage-1 epoch-60 checkpoint (stage-2 init) |
| `vla_reid_encoder.pth` | 102,195,690 | stage-2 encoder export (feature extraction) |
| `vla_reid_cas.pth` | 135,772,354 | stage-2 FCAE+CAS checkpoint (tracker injection) |
| `aflink.pth` | 4,348,705 | AFLink post-linking weight |

Code license: MIT (see the GitHub repo's NOTICE for vendored-component licenses). `sam3_bee24_finetuned.pt` is a derivative of Meta AI's SAM3 checkpoint and follows the SAM license.

If you use the BEE24 benchmark, please cite:

```bibtex
@ARTICLE{10851814,
  author={Cao, Xiaoyan and Zheng, Yiyao and Yao, Yao and Qin, Huapeng and Cao, Xiaoyu and Guo, Shihui},
  journal={IEEE Transactions on Image Processing},
  title={TOPIC: A Parallel Association Paradigm for Multi-Object Tracking Under Complex Motions and Diverse Scenes},
  year={2025},
  volume={34},
  pages={743-758},
  doi={10.1109/TIP.2025.3526066}}
```