Home

Awesome

Multi-Scale Context Aggregation by Dilated Convolutions

Introduction

Properties of dilated convolution are discussed in our ICLR 2016 conference paper. This repository contains the network definitions and the trained models. You can use this code together with vanilla Caffe to segment images using the pre-trained models. If you want to train the models yourself, please check out the document for training.

If you are looking for dilation models with state-of-the-art performance and Python implementation, please check out Dilated Residual Networks.

Citing

If you find the code or the models useful, please cite this paper:

@inproceedings{YuKoltun2016,
	author    = {Fisher Yu and Vladlen Koltun},
	title     = {Multi-Scale Context Aggregation by Dilated Convolutions},
	booktitle = {ICLR},
	year      = {2016},
}

License

The code and models are released under the MIT License (refer to the LICENSE file for details).

Installation

Caffe

Install Caffe and its Python interface. Make sure that the Caffe version is newer than commit 08c5df.

Python

The companion Python script is used to demonstrate the network definition and trained weights.

The required Python packages are numba numpy opencv. Python release from Anaconda is recommended.

In the case of using Anaconda

conda install numba numpy opencv

Running Demo

predict.py is the main script to test the pre-trained models on images. The basic usage is

python predict.py <dataset name> <image path>

Given the dataset name, the script will find the pre-trained model and network definition. We currently support models trained from four datasets: pascal_voc, camvid, kitti, cityscapes. The steps of using the code is listed below:

Training

You are more than welcome to train our model on a new dataset. To do that, please refer to the document for training.

Implementation of Dilated Convolution

Besides Caffe support, dilated convolution is also implemented in other deep learning packages. For example,