Awesome

n-Reference Transfer Learning for Saliency Prediction

n-Reference Transfer Learning enables efficient transfer of knowledge learned from the existing large-scale saliency datasets to a target domain with limited labeled examples. The code is tested under Ubuntu 1804 LTS with Python 3.7.

Requirements

It requires PyTorch 1.2.0+.

conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch

Data

The saliency prediction task requires blurred fixation maps and/or binary fixation maps. Images and corresponding binary fixation maps (i.e., ground truth) can be downloaded via the following links.

• SALICON
• MIT1003
• Webpage saliency dataset (FiWI)
• CAT2000

Note that SALICON and FiWI only provide binary fixation maps. FiWI provides a Matlab script to generate blurred fixation maps. You can either refer to the script to do the preprocessing yourself or download the preprocessed blurred fixation maps via this link.

Core Modules

The reference process is implemented in models.py, named Referencer. TR-Ref is implemented in train_nref.py. TR, FT, and FT|Ref with n samples are implemented in train_nshot.py. Conventional TR is implemented in train.py.

Training

To train a model in various schemes (i.e., TR, TR-Ref, FT, and FT|Ref), run

./train_script.sh

Note that variable split_file indicates the path to the file of the sample splits (i.e., subsets of training/validation/reference samples). If the variable is empty, it will generate ramdon subsets. The splits used in this work can be found in this shared folder.

Citation

If you think our work is interesting or useful, please cite our paper

@InProceedings{Luo_2020_ECCV,
author = {Luo, Yan and Wong, Yongkang and Kankanhalli, Mohan S. and Zhao, Qi},
title = {n-Reference Transfer Learning for Saliency Prediction},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
month = {September},
year = {2020}
}