This is my implementation of the salient object detection method described in
The original implementation was in Matlab and Caffe. I converted it in Python and Tensorflow.
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.
- Linux
- Python 3
- Tensorflow 2
- Unzip the files to a local folder.
- Download the weights
- Donwload the MSO dataset
- Run demo.py.
You can also run the jupyter notebook. If you don't want to install anything, you can execute the notebook with Google Colab.
You can reproduce the result on the MSO dataset reported in the paper, by run benchmark_MSO.py. It will automatically download the MSO dataset and the pre-trained VGG16 model.
The results are the same as the matlab implementation.
To change some configurations, please check get_Param.py.
There is an heuristic window refining process for small objects like in the matlab implementation. Note that this process is not included in the paper or used in the evaluation.