Skeleton-CutMix

This is the repository for the paper Skeleton-CutMix: Mixing Up Skeleton with Probabilistic Bone Exchange for Supervised Domain Adaptation.

You may need to manually change paths for datasets in some places.

Cross-dataset setting

We use NTU-60 and ETRI-Activity3D for the cross-dataset setting. We use 23 action pairs. We evaluate on both NTU $\rightarrow$ ETRI and ETRI $\rightarrow$ NTU settings.

Data preparation

Get skeleton data for both datasets, and follow the readme.txt in data_gen to generate data files. Then you will have *.npy data files and *.pkl label files required by .yaml experiment config files.

Experiments

Run baseline (S+T)

sh rev/common23_new_base.sh <device> <config_idx>

Run Skeleton-CutMix-S

sh rev/common23_new_beta.sh <device> <config_idx>

Run Skeleton-CutMix-W

Prepare weight matrix for part importance sampling (PIS). Follow rev/prepare_pis.txt to run model for each body part, gather classification accuracy scores, and calculate the weight. You may need to try different $T$ depending on different source domains. Or you can directly use our precomputed weight defined in skcutmix_pis_new.py

Then sh rev/common23_new_pis.sh <device> <T>

Acknowledgement

Some of the codes are borrowed from ST-GCN, 2s-AGCN, DAGE. Thanks for their great work!

Contact

Hanchao Liu [liuhc21 at mails.tsinghua.edu.cn]