ARTNet Pytorch

Introduction

Pytorch implementation of ARTNet - Appearance and Relation Network for Video Classification

Paper: http://openaccess.thecvf.com/content_cvpr_2018/html/Wang_Appearance-and-Relation_Networks_for_CVPR_2018_paper.html

Original Caffe implementation: https://github.com/wanglimin/ARTNet

Requirements

The code is written using the following environment. There isn't a strict version requirement, but deviate from the listed versions at your own risk

• python: 3.7.3
• pytorch: 1.2.0
• torchvision: 0.4.0
• matplotlib: 3.1.0
• numpy: 1.16.4
• tqdm: 4.32.1

Training

Preparing Data

The dataset folder should have the following structure

train  
│
└───category1
│   │   
│   │
│   └───video1
│       │   frame001.png
│       │   frame002.png
│       │   ...
│   
└───category2
    │   
    │   ... 

Video frames have to be extracted prior to the training process. This repository, as of now, does not provide means to extract video frames.

Configuration

1. Make a copy of config.ini
2. Edit the configurations as you see fit

Run

python train.py --config [config file path]

Testing

[TODO]

FAQ

Q. Have you tested the code on any standard datasets?

A. As of now, no. I've been having trouble downloading the Kinetic dataset in full, so I've only been able to test the code on a few categories. I'll provide the full testing statistics as soon as possible.