Skip to content
Pseudo-3D Convolutional Residual Networks for Video Representation Learning
C++ Cuda
Branch: master
Clone or download
Latest commit e91f95d Nov 10, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
caffe_add_layers fix linux bugs & change to Caffe engine Oct 17, 2017
prototxt add optical flow network Oct 30, 2017
LICENSE Initial commit Oct 16, 2017
README.md Update README.md Nov 10, 2017

README.md

Pseudo-3D Residual Networks

By Zhaofan Qiu, Ting Yao, Tao Mei.

Microsoft Research Asia (MSRA).

Table of Contents

  1. Introduction
  2. Citation
  3. Implementation
  4. Models
  5. Other Implementation
  6. Contact

Introduction

This repository contains the P3D ResNet models described in the paper "Learning Spatio-Temporal Representation with Pseudo-3D Residual Networks" (http://openaccess.thecvf.com/content_iccv_2017/html/Qiu_Learning_Spatio-Temporal_Representation_ICCV_2017_paper.html). These models are used in ActivityNet 2017 challenge, which won the 1st place in dense-Captioning Events in Videos task and 2rd place in Temporal Action Proposals task.

Citation

If you use these models in your research, please cite:

@inproceedings{qiu2017learning,
  title={Learning Spatio-Temporal Representation with Pseudo-3D Residual Networks},
  author={Qiu, Zhaofan and Yao, Ting and Mei, Tao},
  booktitle={ICCV},
  year={2017}
}

Implementation

  1. We implement the P3D ResNet using our modified Caffe on Windows platform. For fast utilization of our model, here we give the addtional layers used in the network. So you can easily add these layers to your own Caffe branch or Caffe master branch to support P3D ResNet.
  2. In the P3D ResNet, all the blobs are 5D-blobs (num, channels, length, height, width). Some layers in early-version Caffe may only support 4D blobs due to the use of Blob::num(), Blob::channels(), Blob::height() and Blob::width(). You may need to replace these callings with Blob::shape(i).
  3. When training the network, to speedup the network and reduce the memory demand, we use cudnn implementation for conv_layer, relu_layer, bn_layer.
  4. P3D ResNet for ResNext/DenseNet/SENet with P3D convolution and P3D ResNet with lighter weights are in the plan. And our custom Caffe and training/finetuning setting files will be pulic soon.
  5. The mean value for each frame is [104, 117, 123], for each optical flow image is 128. For TVL1 optical flow, we merge the x & y direction grey-level flow image as two-channel image.
  6. For both frame and optical flow, we train the network with 160*160 input resolution as described in our paper. When appling this pre-trianed model or using as feature extractor, larger resolution may get higher performance.

Models

  1. P3D ResNet trained on Sports-1M dataset:

  2. P3D Resnet trained on Kinetics dataset:

Other Implementation

  1. P3D-Pytorch by qijiezhao

Contact

If there is any question, pls feel free to contact me at zhaofanqiu@gmail.com.

You can’t perform that action at this time.