Home

Awesome

SOC-DataAug

This repo contains the data augmentation part used in Salient Objects in Clutter.

You can adapt the settings in config.py for your own trials.

In this document, we investigate three dataset-enhancement strategies (DES), including label smoothing to implicitly emphasize salient boundaries, random image augmentation to adapt saliency models to various scenarios, and self-supervised learning as a regularization strategy to learn from small datasets.

Despite the apparent simplicity of our strategies, we can achieve an average absolute improvement of 1.1% Sα over five existing cutting-edge models:

<img src="materials/table_10.png" alt="table_10" style="zoom:50%;" />

Further, we investigate the contribution of each data-enhancement strategy, and show the performance in Table 11, where we choose CVAE as the base model. “LS”, “RDA”, “SS” represent adding label smoothing strategy, random data augmentation and self-supervised learning to the base model respectively.

<img src="materials/table_11.png" alt="table_11" style="zoom:50%;" />
@article{Fan2021SOC,
title={Salient Objects in Clutter},
author={Fan, Deng-Ping and Zhang, Jing and Xu, Gang and Cheng, Ming-Ming and Shao, Ling},
journal={arXiv preprint arXiv:2105.03053},
year={2021}
}