Convolutional Network In Activity Recognition

This project implement some two-stream convolutional network.

Origin Two-Stream
TSN
DTPP

Data

UCF101

UCF101 contains 101 actions, 13320 video clips.The dataset can be download hereUCF Dataset. About 6.93GB.|

Video -> img

ffmpeg can capture the video's image in one line. Opencv can also do this.

    'ffmpeg -i \"{}\" -vf scale=-1:240 \"{}/image_%05d.jpg\"'.format(video_file_path, dst_directory_path)

details can be find in video_jpg_ucf101_hmdb51.py

img -> flow

FlowNet2.0 is used here to get the flow. Use Docker to finish this part. I use the two job below.
NVIDIA-flownet2-python
lmb

flow -> img

every flo change into two img, u and v.

Transfer Learning

Models

This part include the basebone model in the network.

Four models include here.

bninception
INceptionv4
ResNet
Inceptionv3

caffe to pytorch

caffe to torch

torch to pytorch

spatial_convnet

motion_convnet

fusion

average_fusion and svm_fusion include here.

Reference

These module is based on pytorch.
Pretrained module is based on Cadene
This origin project is based on jerryhuang's project. two-stream-action-recognition