iResNet

This repository contains the code (in CAFFE) for "Learning for Disparity Estimation through Feature Constancy" paper (CVPR 2018 and ROB 2018) by Zhengfa Liang.

Citation

@article{Liang2018Learning,
  title={Learning for Disparity Estimation through Feature Constancy},
  author={Liang, Zhengfa and Feng, Yiliu and Guo, Yulan and Liu, Hengzhu and Chen, Wei and Qiao, Linbo and Zhou, Li and Zhang, Jianfeng},
  booktitle={Computer Vision and Pattern Recognition},
  year={2018},
}

Contents

1. Usage
2. Contacts

Usage

Dependencies

• Ubuntu 16.04
• Python2.7
• Caffe
• CUDNN 5.1
• CUDA 8.0
• Scene Flow
• ETH3D2017
• Kitti2015
• Middlebury2014

Notes:

• You should first install Caffe following the Installation instructions here.

make clean
make all -j 12 tools

• The caffe code in this repository is modiffied from DispNet, which includes the "Correlation1D" layer.

• The FlowWarp layer is from FlowNet 2.0.

• We add RandomCrop layer and DataSwitch layer.

• RandomCrop is used to crop bottom blob to desired width and height, but channel number of this layer is fixed to 7 (left image, right image, and disparity). If the desired width or height is larger than that of bottom blob, we use 128 to fill the first 6 channels, and use NaN to fill the last channel.

layer {  name: "Random_crop_kitti2015"
  type: "RandomCrop"
  bottom: "kitti2015_data"
  top: "kitti2015_cropped_data"
  random_crop_param { target_height: 350  target_width: 694}
}

• DataSwitch is used to randomly select one of the input bottom blobs as output.

layer {  name: "Random_select_datasets"
  type: "DataSwitch"
  bottom: "MiddleBury_cropped_data"
  bottom: "kitti2015_cropped_data"
  bottom: "eth3d_cropped_data"
  top: "curr_data"
}

Data preparation

Download datasets using the instructions from http://www.cvlibs.net:3000/ageiger/rob_devkit. Put the folder "datasets_middlebury2014" under "CAFFE_ROOT/data". The file structure looks like:

+── CAFFE_ROOT
│   +── data
│       +── datasets_middlebury2014
│           +── metadata
│           +── test
│           +── training

For Scene Flow dataset, we only use the FlyingThings3D subset. Please download RGB cleanpass images and its disparity. The file structure looks like:

+── CAFFE_ROOT
│   +── data
│       +── FlyingThings3D_release
│           +── disparity
│           +── frames_cleanpass

Training

2. Enter folder "CAFFE_ROOT/data", and use MATLAB to run the script "reshape_dataset.m"

3. Open terminal, enter folder "CAFFE_ROOT/data", and run the script "make_lmdbs.sh" (replace CAFFE_ROOT first):

sh ./make_lmdbs.sh

Note that, if folder xxxx_lmdb exists, you should first delete this folder, in order to correctly making lmdbs.

4. Enter folder "CAFFE_ROOT/models/ROB_training", and replace CAFFE_ROOT in the xxxx.prototxt under folder "ROB_training". Then run:

python ../train_rob.py 2>&1 | tee rob.log

Evaluattion

Download the pretrained model from [Pretrained Model], and place it in the folder CAFFE_ROOT/models/model. You need to modify CAFFE_ROOT at line 15 in file "test_rob.py". The results for submission will be stored at CAFFE_ROOT/models/submission_results.

  cd models
  python test_rob.py model/iResNet_ROB.caffemodel

Pretrained Model

CVPR 2018

ROB 2018

Contact

[email protected]