RetailAction Challenge weights (team sadbodh)
Trained weights for team sadbodh's entry to the RetailAction Challenge (PRAW workshop @ WACV 2027). Code, Dockerfile and instructions: github.com/sadbodhs/retail-action-challenge.
| File | Size | Contents | Public leaderboard |
|---|---|---|---|
retail-action-weights-v10.tar |
1.0 GB | final submission (v10): 8 pose models, 6 touch-point refiners (fp16); also covers v7 and v8 | 0.9327 |
retail-action-weights-v9.tar |
0.67 GB | v9: the v8 recipe retrained on train + test + validation | 0.9298 |
Each archive unpacks to weights/v*/… inside the repository. From a clone of the GitHub repository, with the
competition zip extracted to /data/extracted:
scripts/download_weights.sh # v10 set, every file checked against weights/SHA256SUMS
RA_DATA=/data/extracted scripts/infer.sh # writes submissions/submission.csv
cmp submissions/submission.csv submissions/v10_v2b.csv # no output: identical to the submitted file
Requirements: Linux, Docker with the NVIDIA Container Toolkit, an NVIDIA GPU with ≥ 16 GB memory and compute capability 7.5 or newer, NVIDIA driver 580 or newer, and about 75 GB free disk. On one RTX 3090 the run takes about 15 minutes and reproduces the submitted file byte for byte.
SHA256SUMS in this repository lists the checksums of the two archives.
Licence
CC-BY-NC-4.0. The two ConvNeXt-V2-B refiners (weights/v10/refinerz_v2b.pt, weights/v10/refinerf_v2b.pt) are
fine-tuned from timm's convnextv2_base.fcmae_ft_in22k_in1k, which is CC-BY-NC-4.0, so these weights are
non-commercial. The ConvNeXt-S and ConvNeXt-B refiners start from torchvision ImageNet-1k weights, and the pose models
are trained from scratch. All weights are trained on the
RetailAction dataset, whose own terms also apply.
The code in the GitHub repository is MIT-licensed.