juuuncha/praw-retailaction-dxdshs-weights

DX_DSHS — PRAW RetailAction Challenge (WACV 2027) submission weights

Fine-tuned checkpoints used by https://github.com/dxlabcolab-arch/praw-retailaction-dxdshs (inference code + Docker). Downloaded automatically by that repo's download_weights.sh; MD5SUMS lists every file.

file	what
stage1_1.pt … stage1_4.pt	stage 1 (V-JEPA 2 + FACT-DETR), four models of the same architecture (ViT-L ×3, ViT-g ×1), fused at inference
refiner_1.pt … refiner_6.pt	stage 2 point refiners (V-JEPA 2.1 ViT-L), all labelled data, six random seeds. Final submission 1 (ens6_vitl) = mean of all six; final submission 2 (single_vitl) = refiner_1.pt alone; ens4_vitl = 1, 2, 3, 6
expected_<preset>.csv	the submitted prediction files, for byte-level verification

Backbones are fine-tuned from Meta's V-JEPA 2 (MIT) and V-JEPA 2.1 (see Meta's license); they inherit those terms.