Installation

Our experimental environment:

Ubuntu 18.04.5 Python 3.9.6, PyTorch 1.8.1.

During training we use 2 NVIDIA GeForce RTX 3090 with cuda11.1 (About 48G memory needed).

1.Basic Usage

1. Create a new conda environment and activate the environment.

   conda create --n IOD python=3.9
   conda activate IOD
   

2. Install pytorch1.8.1:

   conda install pytorch==1.8.1 torchvision cudatoolkit -c pytorch
   

   The torch version is not supposed to be 1.8.1, it can be changed to any other version, only the cuda version is required to be same with CUDA environment(/usr/local/cuda) for Pytorch Correlation extension.

   
   

3. Clone this repo (${MOC_ROOT} is the path to clone):

   git clone https://github.com/CalayZhou/IOD-Video.git 
   

4. Install the requirements

   pip install -r pip-list.txt
   

2.Additional Usage

1. cuda_shift for TIN

   cd ./network/twod_models/cuda_shift
   bash make.sh
   

   You can refer to TIN for detail.

2. Pytorch-Correlation-extension for MSNet

   git clone https://github.com/ClementPinard/Pytorch-Correlation-extension.git
   python setup.py install 
   

   The cuda version of torch should be same with CUDA environment( /usr/local/cuda), otherwise it may cause issue as mentioned in #80. If GPUs are NVIDIA 30 series, the version can be changed to the branch fix_1.7. You can refer to MSNet and Pytorch Correlation extension for more information.

The basic usage can satisfy most spatio-temporal backbones except TIN and MSNet. You can comment TINresnet and MSresnet in STA_Framework.py for fast implementation.