DeconvNet: Learning Deconvolution Network for Semantic Segmentation

Created by Hyeonwoo Noh, Seunghoon Hong and Bohyung Han at POSTECH

Acknowledgements: Thanks to Yangqing Jia and the BVLC team for creating Caffe.

Introduction

DeconvNet is state-of-the-art semantic segmentation system that combines bottom-up region proposals with multi-layer decovolution network.

Detailed description of the system will be provided by our technical report [arXiv tech report] http://arxiv.org/abs/1505.04366

Citation

If you're using this code in a publication, please cite our papers.

@article{noh2015learning,
  title={Learning Deconvolution Network for Semantic Segmentation},
  author={Noh, Hyeonwoo and Hong, Seunghoon and Han, Bohyung},
  journal={arXiv preprint arXiv:1505.04366},
  year={2015}
}

Pre-trained Model

If you need model definition and pre-trained model only, you can download them from following location: 0. caffe for DeconvNet: https://github.com/HyeonwooNoh/caffe 0. DeconvNet model definition: http://cvlab.postech.ac.kr/research/deconvnet/model/DeconvNet/DeconvNet_inference_deploy.prototxt 0. Pre-trained DeconvNet weight: http://cvlab.postech.ac.kr/research/deconvnet/model/DeconvNet/DeconvNet_trainval_inference.caffemodel

Licence

This software is being made available for research purpose only. Check LICENSE file for details.

System Requirements

This software is tested on Ubuntu 14.04 LTS (64bit).

Prerequisites 0. MATLAB (tested with 2014b on 64-bit Linux) 0. prerequisites for caffe(http://caffe.berkeleyvision.org/installation.html#prequequisites)

Installing DeconvNet

By running "setup.sh" you can download all the necessary file for training and inference include: 0. caffe: you need modified version of caffe which support DeconvNet - https://github.com/HyeonwooNoh/caffe.git 0. data: data used for training stage 1 and 2 0. model: caffemodel of trained DeconvNet and other caffemodels required for training

Training DeconvNet

Training scripts are included in ./training/ directory

To train DeconvNet you can simply run following scripts in order: 0. 001_start_train.sh : script for first stage training 0. 002_start_train.sh : script for second stage training 0. 003_start_make_bn_layer_testable : script converting trained DeconvNet with bn layer to inference mode

Inference EDeconvNet+CRF

Run run_demo.m to reproduce EDeconvNet+CRF results on VOC2012 test data.

This script will generated EDeconvNet+CRF results through following steps: 0. run FCN-8s and cache the score [cache_FCN8s_results.m] 0. generate DeconvNet score and apply ensemble with FCN-8s score, post processing with densecrf [generate_EDeconvNet_CRF_results.m]

EDeconvNet+CRF obtains 72.5 mean I/U on PASCAL VOC 2012 Test

External dependencies [can be downloaded by running "setup.sh" script] 0. FCN-8s model and weight file [https://github.com/BVLC/caffe/wiki/Model-Zoo] 0. densecrf with matlab wrapper [https://github.com/johannesu/meanfield-matlab.git] 0. cached proposal bounding boxes extracted with edgebox object proposal [https://github.com/pdollar/edges]