codename1995 / ASD-MF

This repository contains Keras implementation of our atypical visual saliency prediction model.

SALIENCY PREDICTION VIA MULTI-LEVEL FEATURES AND DEEP SUPERVISION FOR CHILDREN WITH AUTISM SPECTRUM DISORDER

This repository contains Keras implementation of our atypical visual saliency prediction model.

Cite

Please cite with the following Bibtex code:

@inproceedings{wei2019saliency,
  title={Saliency prediction via multi-level features and deep supervision for children with autism spectrum disorder},
  author={Wei, Weijie and Liu, Zhi and Huang, Lijin and Nebout, Alexis and Le Meur, Olivier},
  booktitle={2019 IEEE International Conference on Multimedia \& Expo Workshops (ICMEW)},
  pages={621--624},
  year={2019},
  organization={IEEE}
}

Pretrained weight on Saliency4ASD

Google Drive

Training

Train model from scratch

$ python train.py --train_set_path path/to/training/set --val_set_path path/to/validation/set 

For training model based on our pretrained weight, please download the weight file and put it into weights/.

$ python train.py --train_set_path path/to/training/set --val_set_path path/to/validation/set --model_path weights/weights--1.4651.pkl

The dataset directory structure should be

└── Set  
    ├── Images  
    │   ├── 1.png  
    │   └── ...
    ├── FixMaps  
    │   ├── 1.png  
    │   └── ...
    ├── maps  
        ├── 1.mat  
        └── ...

Note: We convert the *_f.png files in Saliency4ASD\TrainingDataset\AdditionalData\ASD_FixPts\ to MAT file by following code:

% Matlab Code
im = imread('1_f.png');
save('1.mat', 'im');

Testing

Clone this repository and download the pretrained weights.

Then just run the code using

$ python test.py --images_path path/to/test/images --results_path path/to/results --model_path path/to/saved/models

This will generate saliency maps for all images in the images directory and save them in results directory

Requirements:

cuda 8.0
cudnn 5.1
python 3.5
keras 2.2.2
theano 0.9.0
opencv 3.1.0
matplotlib 2.0.2

Update:

2020/12/17

The original Saliency4ASD only contains FixPts in PNG format. We provide a simple code to convert the PNG file to MAT file for easy-using of our model.

Acknowledgement

The code is heavily inspired by the following project:

1. SAM : https://github.com/marcellacornia/sam

Thanks for their contributions.

Contact

If you have any questions, please contact me at [e-mail] or my supervisor Prof. Zhi Liu at [e-mail].

License

This code is distributed under MIT LICENSE.