grseb9s/hca_saliency_codes

Codes for HCA (Hierarchical Cellular Automata for Visual Saliency) accepted by IJCV 2018.

Hierarchical Cellular Automata for Visual Saliency

Introduction

HCA is a temporally evolving model to intelligently detect salient objects. This package contains the source codes to reproduce the experimental results of HCA. The source code is mainly written in MATLAB.

Publication

Our paper "Hierarchical Cellular Automata for Visual Saliency" has been accepted for publication in Iternational Journal of Computer Vision (IJCV), 2018. [Online-Version]

License

This code is released under the MIT License (refer to the LICENSE file for details).

Contents

1. Requirements: software
2. Requirements: hardware
3. Basic installation
4. Demo
5. Pre-computed saliency maps
6. Visual comparison with state-of-the-art methods

Requirements: software

1. Requirements for MatConvNet (see: MatConvNet installation instruction).

2. MATLAB.

3. [optional] CUDA (we use CPU to compute FCN features, if you want to use GPU, please compile MatConvNet with CUDA enabled).

4. Supported OS: the source code was tested on 64-bit Windows OS, it used SLIC to pre-process the images into super-pixels. Here we used the mex file in Windows OS, so the HCA code may not worked on Linux OS for now.

Requirements: hardware

If you compile MatConvNet with CUDA supported, a GPU with at least 3G of memory suffices.

Installation (sufficient for the demo)

1. Clone the HCA repository

git clone https://github.com/ArcherFMY/HCA_saliency_codes.git

2. cd to the root directory of HCA (we will call the directory HCA_ROOT), use MATLAB to run compile_matconvnet.m.

3. Download the pre-trained FCN-32s models from here. Then put it under $HCA_ROOT/matconvnet-1.0-beta19/Data/ folder with name pascal-fcn32s-dag.mat.

note

Here we just compiled the MatConvNet with CPU. Users could compile with GPU supported yourself.

Demo

To run the demo, simply run $HCA_ROOT/runme.m with MATLAB. Saliency maps will be saved in $HCA_ROOT/saliencmaps/ folder.

Precomputed saliency maps

We provided pre-computed saliency maps for convenience.

Included Datasets: ECSSD, HKU-IS, DUT-OMRON, PASCAL-S and MSRA5000.

pre-computed saliency maps

Visualization

FAQ

Q: Error when running extract_fcn_im_features at line(23)

A: We modified the matconvnet-1.0-beta19/matlab/+dagnn/@DagNN/eval.m to allow users to extract features from every layers (conv, pool, relu). So if you are using your own matconvnet, please modify the corresponding .m file as we did.