http://cvlab.cse.msu.edu/pdfs/Zhang_Tran_Yin_Atoum_Liu_Wan_Wang_CVPR2019.pdf

• Linux/ Windows
• Anaconda 3 - Python 3.7 version

Raw_Dataset http://cvlab.cse.msu.edu/frontal-view-gaitfvg-database.html

• Download codes

git clone https://github.com/ZiyuanTonyZhang/GaitNet.git

cd GaitNet/

• Install required libs

pip install -r requirements.txt

• OR if you use Anaconda

• Download Dataset cropped images

Unzip to Data/ , do not change the three folder names

The final folder structure should look like:

• Data
  • SEG-S1
  • SEG-S2
  • SEG-S3
• GaitNet
  • train.py
  • runs
  • ...

Since the code is set up with orginal papers defaut hyperparameters, simply run with:

python train.py

You will be asked which GPU to use, enter 0 if you have only one GPU. If you have multiple GPUs, check their index with

nvidia-smi

After running train.py, run TensorboardX to visulize training loss curves and synthesized results:

tensorboard --logdir runs

MRCNN: TORCHVISION

IMAGE PROCESSING: PIL, TORCHVISION

1. MASK R-CNN
   1. torchvision
   2. faster (11-12FPS on 1080P)
   3. threading data loader
   4. more effective algorithm to remove redundant data(out of frame)
2. CUDNN
   1. 3 times fater for training