This is a repository containing the implementation of our ACM MM 2023 paper UmURL.
Use the following instructions to create the corresponding conda environment.
conda create -n umurl python=3.9 anaconda
conda activate umurl
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 -c pytorch
pip3 install tensorboard
- Download raw NTU-RGB+D 60 and 120 skeleton data and save to ./data folder.
- data/
- nturgbd_raw/
- nturgb+d_skeletons/
...
- samples_with_missing_skeletons.txt
- nturgbd_raw_120/
- nturgb+d_skeletons/
...
- samples_with_missing_skeletons.txt
- Preprocess data with
data_gen/ntu_gendata.py.
cd data_gen
python ntu_gendata.py
Our implementation is based on automatic mixed precision (amp) and Distributed Data Parallel (DDP).
- Run the following script for pretraining. It will save the checkpoints to
./checkpoints/ntu60_xsub.
# Here is an example
./script_pretrain.sh
The default setting is to pre-train with NTU-60 x-sub protocol. Please refer to the bash scripts for other pretraining configurations.
- Task1: Skeleton-based action recognition. Train a linear classifier on the pre-trained encoder. Multi-modal representations are utilized here.
# ./script_action_recognition.sh $dir $dataset $protocol
./script_action_recognition.sh ntu60_xsub ntu60 cross_subject
$dir is the folder where the checkpoints are saved.
$dataset is the dataset to use for unsupervised pre-training (ntu60 or ntu120).
$protocol means training protocol (cross_subject/cross_view for ntu60, and cross_subject/cross_setup for ntu120).
It will automatically evaluate the checkpoint from the final epoch obtained from the aforementioned pre-training. Please refer to the bash scripts for other pre-training configurations.
- Task2: Skeleton-based action retrieval. Apply a KNN classifier on the pre-trained encoder. It's similar to action recognition, here is an example.
# ./script_action__retrieval.sh $dir $dataset $protocol
./script_action_retrieval.sh ntu60_xsub ntu60 cross_subject
$dir is the folder where the checkpoints are saved.
$dataset is the dataset to use for unsupervised pre-training (ntu60 or ntu120).
$protocol means training protocol (cross_subject/cross_view for ntu60, and cross_subject/cross_setup for ntu120).
We release several pre-trained models: Google Drive
Expected performance on skeleton-based action recognition and skeleton-based action retrieval:
| Task | NTU 60 x-sub (%) | NTU 60 x-view (%) | NTU 120 x-sub (%) | NTU 120 x-setup (%) |
|---|---|---|---|---|
| action recognition | 84.2 | 90.9 | 75.2 | 76.3 |
| action retrieval | 72.0 | 88.9 | 59.5 | 62.2 |
t-SNE visualization of the learned multi-modal action representations obtained by (a) simple baseline and (b) our proposed UmURL on NTU-60. 10 classes from the x-view testing set are randomly selected for visualization. Dots with the same color indicate actions belonging to the same class.
If you find this repository useful, please consider citing our paper:
@inproceedings{sun2023unified,
title={Unified Multi-modal Unsupervised Representation Learning for Skeleton-based Action Understanding},
author={Sun, Shengkai and Liu, Daizong and Dong, Jianfeng and Qu, Xiaoye and Gao, Junyu and Yang, Xun and Wang, Xun and Wang, Meng},
booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
pages={2973--2984},
year={2023}
}
This work was supported by the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No.2023C01212), Public Welfare Technology Research Project of Zhejiang Province (No. LGF21F020010), National Natural Science Foundation of China (No. 61976188, 62272435, and U22A2094), Young Elite Scientists Sponsorship Program by CAST (No. 2022QNRC001), the open research fund of The State Key Laboratory of Multimodal Artificial Intelligence Systems, and the Fundamental Research Funds for the Provincial Universities of Zhejiang.