One-stage Progressive Dichotomous Segmentation
Jing Zhu (Samsung Research America),* Karim Ahmed (Samsung Research America), Wenbo Li (Samsung Research America), Yilin Shen (Samsung Research America), Hongxia Jin (Samsung Research America)
The 34th British Machine Vision Conference
Abstract
Dichotomous segmentation is a challenging task that involves recognizing foreground objects in high-resolution images with varying characteristics. Existing methods often miss important details of the object or require a long processing time due to multi-stage process. In this paper, we propose a one-stage effective model that can distinguish objects in dichotomous segmentation with low computation cost. Unlike most methods that use two separate branches to first obtain coarse results from low-resolution images and then refine them with the high-resolution information, our method can directly process high-resolution inputs with simple operations. We introduce convolutional attentions into the feature extractor to effectively capture multi-scale features. These features are then used to generate high-quality results with a specifically designed progressive decoder. The experimental results demonstrate that our method achieves superior performance on the DIS5K dichotomous segmentation dataset with fewer model parameters and computational operations.Video
Citation
@inproceedings{Zhu_2023_BMVC,
author = {Jing Zhu and Karim Ahmed and Wenbo Li and Yilin Shen and Hongxia Jin},
title = {One-stage Progressive Dichotomous Segmentation},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0077.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by