pytorch implement of this paper:https://arxiv.org/abs/1807.11176 (ECCV 2018)

Implement by:

Tim Ren, Harrison Huang

• MMD-NCA Loss
• Layer Normalization LSTM
• Self-Attention
• Training
• Improve Dataloader

Instead of a Bi-direction Layer Normalization LSTM, we use a non-normalizaiton bi-direction GRU version.

And for now, the dataloader may use a large memory of your cpu.

If there is any problem, make the parameter: num_MMD_NCA_Groups of "MMD_NCA_Dataset" smaller.

I clean the dance dataset of https://arxiv.org/abs/1801.07388

The cleaned dataset is provided here (update the old link):

https://drive.google.com/file/d/1-H0ywex6KhA68MPiU7i5Apde_MzXg_vK/view?usp=sharing

The dataset contains 16 classes of dance. It contain 51858 sequence. The key of the json file is "0","1",.....,"15" Each key contains: ( _ , 50, 2, 17) pose. 2 is channel, 17 is pose coordinates as coco format. And each pose is normalized.

cd Human-Motion-Analysis-with-Deep-Metric-Learning
mkdir log
mkdir dataset

Download the dataset I provided above, put it in the folder "dataset". It is suggested to split it by yourself, for the dataset is too large.

Note: if you split the data, you need to change line 247 in train.py.

And run:

python train.py

If have any question, feel free to connect me by email: [email protected]