---
license: cc-by-nc-4.0
library_name: timm
tags:
- pedestrian-attribute-recognition
- person-retrieval
- upar
---

# UPAR Challenge 2027 attribute models (team angadbajwa)

Trained weights for our UPAR Challenge 2027 entries (WACV 2027 Real-World Surveillance workshop):
Track 2, attribute-based person retrieval (69.11 mADM on the test set), and Track 1, pedestrian
attribute recognition (79.55).

Code, training recipe and instructions: https://github.com/angadbajwa23/upar2027

| File | Backbone (timm name) | Used in |
|---|---|---|
| `eva02L_lpft_ep6.pt` | `eva02_large_patch14_224.mim_m38m`, LP-FT | Track 1, Track 2 |
| `clipL_metaclip_ep4.pt` | `vit_large_patch14_clip_224.metaclip_2pt5b` | Track 1, Track 2 |
| `clipL_openai_ep4.pt` | `vit_large_patch14_clip_224.openai` | Track 2 |

Each file is a checkpoint written by `train.py` (fp16 EMA state dict under `state_dict`) for the
model in `model.py`: backbone, dropout and a linear layer with 40 sigmoid outputs, at input size 252×126.
`make_submission.py` in the code repository turns them into the Codabench submissions.

SHA-256 checksums are in `weights.sha256` in the code repository.

## License

Non-commercial research use only. The models are trained on the UPAR annotations (CC BY-NC-SA) and
the PA100k, PETA and Market-1501 images, and one of them starts from MetaCLIP weights (CC BY-NC 4.0).