DD-Net

(A Double-feature Double-motion Network)

JHMDB

SHREC-14

Update

Updating code to current Keras environment.

Thanks for @muxizju, @pengfeiZhao1993, and @YLTsai0609 helped to fix bugs in this code. After fixing bugs, the performance is further improved - JHMDB 82-83% and SHREC-14 96-97 %.

Thanks for Nightwatch, who made a pytorch version DD-Net. You can check it at Pytorch DD-Net

1.About this code

A lightweight network for body/hand action recognition, implemented by keras tensorflow backend. It also could be the simplest tutorial code to start skeleton-based action recognition.

2.How to use this code

(1) clone DD-Net

git clone https://github.com/fandulu/DD-Net.git

(2) I just noticed that the environment may not be available due to TensorFlow updating, so it is better to check the setting in and install the currently available environment

(3) go to the folder of JHMDB or SHREC to play with ipython notebooks.

Note: You can download the raw data and use our code to preprocess them, or, directly use our preprocessed data under /data.

JHMDB raw data download link:   http://jhmdb.is.tue.mpg.de/challenge/JHMDB/datasets
SHREC raw data download link:   http://www-rech.telecom-lille.fr/shrec2017-hand/

3.Problems this code try to alleviate

4.Performance

No. parameters	SHREC-14	SHREC-28
1.82 M	94.6	91.9
0.15 M	91.8	90.0

No. parameters	JHMDB
1.82 M	77.2
0.50 M	73.7

Note: if you want to test the speed, please try to run the model.predict() at leat twice and do not take the speed of first run, the model initialization takes extra time.

5.Citation

If you find this code is helpful, thanks for citing our work as,

@inproceedings{yang2019ddnet,
  title={Make Skeleton-based Action Recognition Model Smaller, Faster and Better},
  author={Fan Yang, Sakriani Sakti, Yang Wu, and Satoshi Nakamura},
  booktitle={ACM International Conference on Multimedia in Asia},
  year={2019}
}

Hey, come take a look