Salient object detection with image-level binary supervision. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Pengjie; Liu, Yuxuan; Cao, Ying; Yang, Xin; Luo, Yu; Lu, Huchuan; Liang, Zijian; Lau, Rynson W.H.
• Is Part Of: Pattern recognition. Volume 129(2022)
• Publication Date: 2022-09
• Subjects: Weak supervision -- Salient object detection -- Binary labels; 00-01 -- 99-00; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108782
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a learning framework for salient object detection, which is trained only on image-level binary labels but yields comparable performance to state-of-the-art weakly-supervised methods. We propose a target label hallucination method, which can synthesize quality pseudo ground truth saliency maps from just binary labels. We have conducted extensive experimental evaluation to demonstrate the effectiveness of the proposed weakly-supervised method. Abstract: Recent deep learning based salient object detection (SOD) methods have achieved impressive performance. However, while fully-supervised methods require a large amount of labeled data, weakly-supervised methods still require a considerable human effort. To address this problem, we propose a novel weakly-supervised method for salient object detection based on only binary image tags, which are much cheaper to collect. Our basic idea is to construct a dataset of images that are labeled as either salient (with salient objects) or non-salient (without salient objects), and leverage such binary labels as supervision to learn a salient object detector based on existing unsupervised methods. In particular, we propose a target saliency map hallucinator, which can synthesize pseudo ground truth saliency maps for the salient images in the training data solely from binary labels. We can then use the pseudo ground truth labels to train a salient object detector. Experimental results show that our method performs comparably to the state-of-the-art weakly-supervised methods, but requires considerably less human supervision.