Carridevy/UPAR2027-ninhhh1

UPAR Challenge 2027 weights (team ninhhh1)

Weights for our Track 1 and Track 2 entries in the UPAR Challenge 2027 (RWS workshop at WACV 2027). Code and instructions: https://github.com/CreDevy1/UPAR2027-ninhhh1 Paper: https://openreview.net/forum?id=TvG118JOn1 (submission 65)

Final entries:

File	Content
submissions/t1_ens2_hr_tv_pmcd_20261021.zip	Track 1 package, Codabench submission 1011642 (challenge average 0.7902)
submissions/t2_v5a_hr_fp32.zip	Track 2 package, Codabench submission 1011654 (mADM 0.7036)
runs_hr_tv/allsource_pel_hr_lc_m1_s42/epoch6.pt	PE-Core L/14, 336x336, fine-tuned on TRAIN+VAL (both final entries)
runs_hr_tv/allsource_sl2l_hr_lc_m1_s42/epoch6.pt	SigLIP 2 L/16, 384x384, fine-tuned on TRAIN+VAL (both final entries)

Earlier entries:

File	Content
submissions/t1_pel_hr_tv_A1_20261021.zip	Track 1 package, Codabench submission 1007246 (challenge average 0.7906)
submissions/t2_v4b.zip	Track 2 package, Codabench submission 1004665 (mADM 0.6979)
runs_tv/allsource_pel_lc_m1_s42/epoch6.pt	PE-Core L/14, 224x112, fine-tuned on TRAIN+VAL
runs_tv/allsource_sl2l_lc_m1_s42/epoch6.pt	SigLIP 2 L/16, 256x128, fine-tuned on TRAIN+VAL

The zips are the exact files submitted to Codabench. They run with torch, torchvision, numpy and Pillow only. SHA-256 checksums are in tools/download_weights.py of the code repository.

Base models: facebook/PE-Core-L14-336, google/siglip2-large-patch16-384 and google/siglip2-large-patch16-256, all Apache 2.0.