tannguyenbmt47/upar2027-track2

UPAR Challenge 2027, Track 2 (attribute-based person retrieval): trained weights

Trained weights of our Track 2 leaderboard entry v13 (mADM 0.6817 on the hidden test set) to the UPAR Challenge 2027 (Real-World Surveillance workshop, WACV 2027). Code, training scripts and instructions: https://github.com/tannguyenbmt47/upar2027-track2 . The Codabench zip is rebuilt from these weights with submission/build_v13.py.

File	What it is
student.pt	model 1: bf16 state dict of DINOv3 ViT-L/16 (176 x 96) with a query head, distilled for 20 epochs from DINOv3 ViT-L/16 and CLIP ViT-L/14 teachers (0.2 label loss + 0.8 distillation, half of the batches mixup / CutMix), seed 2
hp.pt	model 2: bf16 state dict of DINOv3 ViT-H+/16 (224 x 112) with a query head, BCE + 0.5 group cross-entropy, EMA weights after 3 epochs, seed 1
m1.pt	model 3: bf16 state dict of OpenAI CLIP ViT-L/14 (224 x 112) with a query head and the CLIP text score, same training as hp.pt, then WiSE-FT 0.7

All three models are trained on the train + val splits of PA100k, PETA and Market-1501.

Download:

hf download tannguyenbmt47/upar2027-track2 --local-dir weights      # huggingface_hub >= 1.0

The weights are fine-tuned from DINOv3 (DINOv3 License) and OpenAI CLIP (MIT License) and keep the terms of those pretrained weights. The code is released under Apache 2.0.