jimmie33/sod

A Method for Unconstrained Salient Object Detection

Unconstrained Salient Object Detection

This is an implementation of the salient object detection method described in

Jianming Zhang, Stan Sclaroff, Zhe Lin, Xiaohui Shen, Brian Price and Radomír Mech. "Unconstrained Salient Object Detection via Proposal Subset Optimization." CVPR, 2016.

This method aims at producing a highly compact set of detection windows for salient objects in uncontrained images, which may or may not contain salient objects. Please cite the above paper if you find this work useful.

Prerequisites

1. Linux
2. Matlab
3. Caffe & Matcaffe (We use the official master branch downloaded on 4/1/2016. Previous versions may not be compatible.)

Quick Start

1. Unzip the files to a local folder (denoted as root_folder).
2. Enter the root_folder and modify the Matcaffe path in setup.m.
3. In Matlab, run setup.m and it will automatically download the pre-trained GoogleNet model.
4. Run demo.m.

Evaluation

You can reproduce the result on the MSO dataset reported in the paper, by run benchmarkMSO.m. It will automatically download the MSO dataset and the pre-trained VGG16 model.

Miscs

To change CNN models or other configurations, please check getParam.m.

In the demo, we use the pre-trained GoogleNet, which is faster and slightly better than the VGG16 model used in our paper. We have also added a heuristic window refining process for small objects. Note that this process is not included in our paper or used in our evaluation.

sod's Issues

How to leverage the model on my own dataset

Hi, we are interested in your work. Our work needs to leverage your method to generate salient area/bbox. Can you guide me in detail how to use your codes on my dataset? I am really confused because I mainly use python to write my codes now. Many thanks!

How to train the vgg16 or Imagenet model

I want to train the VGG16_SOD_finetune.caffemodel by myself, but I don't know how to write the solver prototxt.
Can you give me a detailed information about the solver prototxt to train and get the VGG16_SOD_finetune model?