Home

Awesome

Shuffle and Learn (Shuffle Tuple)

Created by Ishan Misra

Based on the ECCV 2016 Paper - "Shuffle and Learn: Unsupervised Learning using Temporal Order Verification" link to paper.

This codebase contains the model and training data from our paper.

Introduction

Our code base is a mix of Python and C++ and uses the Caffe framework. Design decisions and some code is derived from the Fast-RCNN codebase by Ross Girshick.

Citing

If you find our code useful in your research, please consider citing:

@inproceedings{misra2016unsupervised,
  title={{Shuffle and Learn: Unsupervised Learning using Temporal Order Verification}},
  author={Misra, Ishan and Zitnick, C. Lawrence and Hebert, Martial},
  booktitle={ECCV},
  year={2016}
}

Benchmark Results

We summarize the results of finetuning our method here (details in the paper).

Action Recognition

| Dataset | Accuracy (split 1) | Accuracy (mean over splits) :--- | :--- | :--- | :--- UCF101 | 50.9 | 50.2 HMDB51 | 19.8 | 18.1

Pascal Action Classification (VOC2012): Coming soon

Pose estimation

Object Detection

We initialize conv1-5 using our unsupervised pre-training. We initialize fc6-8 randomly. We then follow the procedure from Krahenbuhl et al., 2016 to rescale our network and finetune all layers using their hyperparameters.

Surface Normal Prediction

Contents

  1. Requirements: software
  2. Models and Training Data
  3. Usage
  4. Utils

Requirements: software

  1. Requirements for Caffe and pycaffe (see: Caffe installation instructions)

Note: Caffe must be built with support for Python layers and OpenCV.

# In your Makefile.config, make sure to have this line uncommented
WITH_PYTHON_LAYER := 1
USE_OPENCV := 1

You can download a compatible fork of Caffe from here. Note that since our model requires Batch Normalization, you will need to have a fairly recent fork of caffe.

Models and Training Data

  1. Our model trained on tuples from UCF101 (train split 1, without using action labels) can be downloaded here.

  2. The tuples used for training our model can be downloaded as a zipped text file here. Each line of the file train01_image_keys.txt defines a tuple of three frames. The corresponding file train01_image_labs.txt has a binary label indicating whether the tuple is in the correct or incorrect order.

  3. Using the training tuples requires you to have the raw videos from the UCF101 dataset (link to videos). We extract frames from the videos and resize them such that the max dimension is 340 pixels. You can use ffmpeg to extract the frames. Example command: ffmpeg -i <video_name> -qscale 1 -f image2 <video_sub_name>/<video_sub_name>_%06d.jpg, where video_sub_name is the name of the raw video without the file extension.

Usage

  1. Once you have downloaded and formatted the UCF101 videos, you can use the networks/tuple_train.prototxt file to train your network. The only complicated part in the network definition is the data layer, which reads a tuple and a label. The data layer source file is in the python_layers subdirectory. Make sure to add this to your PYTHONPATH.
  2. Training for Action Recognition: We used the codebase from here
  3. Training for Pose Estimation: We used the codebase from here. Since this code does not use caffe for training a network, I have included a experimental data layer for caffe in python_layers/pose_data_layer.py

Utils

This repo also includes a bunch of utilities I used for training and debugging my models