I3D models

Overview

This repository contains reproducing code reported in the paper "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset" by Joao Carreira and Andrew Zisserman.

Running the code

Setup

Firstly, clone this repository using

$ git clone https://github.com/dythebs/i3d_reproducing.git

Then, download Dataset and List using

$ curl -O http://crcv.ucf.edu/data/UCF101/UCF101.rar
$ curl -O http://crcv.ucf.edu/data/UCF101/UCF101TrainTestSplits-RecognitionTask.zip

Data Preprocessing

Mov uncompressed files to /data

$ python avi2train.py

Train

$ python inception_v1_3d_train.py

Repo Structure

.
├── ckpt
│   ├── inception_v1_2016_08_28.tar.gz
│   └── inception_v1.ckpt
├── data
│   ├── UCF-101
│   ├── UCF-101_flow_np_train
│   ├── UCF-101_rgb_np_train
│   ├── avi2train.py
│   └── inception_v1.ckpt
├── ckpt_util.py
├── inception_v1_3d.py
├── inception_v1_3d_train.py
└── input_data.py

Reference

kinetics-i3d

Thank Joao Carreira for offering me help.