acewjh/pytorch-video-recognition

PyTorch implemented C3D, R3D, R2Plus1D models for video activity recognition.

pytorch-video-recognition

Introduction

This repo contains several models for video action recognition, including C3D, R2Plus1D, R3D, inplemented using PyTorch (0.4.0). Currently, we train these models on UCF101 and HMDB51 datasets. More models and datasets will be available soon!

Note: An interesting online web game based on C3D model is in here.

Installation

The code was tested with Anaconda and Python 3.5. After installing the Anaconda environment:

0. Clone the repo:

   git clone https://github.com/jfzhang95/pytorch-video-recognition.git
   cd pytorch-video-recognition

1. Install dependencies:

   For PyTorch dependency, see pytorch.org for more details.

   For custom dependencies:

   conda install opencv
   pip install tqdm scikit-learn tensorboardX

2. Download pretrained model from BaiduYun or GoogleDrive. Currently only support pretrained model for C3D.

3. Configure your dataset and pretrained model path in mypath.py.

4. You can choose different models and datasets in train.py.

   To train the model, please do:

   python train.py

Datasets:

I used two different datasets: UCF101 and HMDB.

Dataset directory tree is shown below

• UCF101 Make sure to put the files as the following structure:

  UCF-101
  ├── ApplyEyeMakeup
  │   ├── v_ApplyEyeMakeup_g01_c01.avi
  │   └── ...
  ├── ApplyLipstick
  │   ├── v_ApplyLipstick_g01_c01.avi
  │   └── ...
  └── Archery
  │   ├── v_Archery_g01_c01.avi
  │   └── ...
  

After pre-processing, the output dir's structure is as follows:

ucf101
├── ApplyEyeMakeup
│   ├── v_ApplyEyeMakeup_g01_c01
│   │   ├── 00001.jpg
│   │   └── ...
│   └── ...
├── ApplyLipstick
│   ├── v_ApplyLipstick_g01_c01
│   │   ├── 00001.jpg
│   │   └── ...
│   └── ...
└── Archery
│   ├── v_Archery_g01_c01
│   │   ├── 00001.jpg
│   │   └── ...
│   └── ...

Note: HMDB dataset's directory tree is similar to UCF101 dataset's.

Experiments

These models were trained in machine with NVIDIA TITAN X 12gb GPU. Note that I splited train/val/test data for each dataset using sklearn. If you want to train models using official train/val/test data, you can look in dataset.py, and modify it to your needs.

Currently, I only train C3D model in UCF and HMDB datasets. The train/val/test accuracy and loss curves for each experiment are shown below:

• UCF101

• HMDB51

Experiments for other models will be updated soon ...