Low-Light Salient Object Detection Meets the Small Size

Wang, Shiqin1; Xu, Xin1; Chen, Haoyang2; Jiang, Kui3; Wang, Zheng2; Tang, Ke4

2024-08-01

10.1109/TETCI.2024.3442819

IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE   影响因子和分区

2471-285X

Low-light Small Salient Object Detection (LS-SOD) focuses on small salient objects in low light, a crucial and realistic problem for nighttime automatic driving and surveillance. However, relatively few efforts have been put toward LS-SOD due to the challenge of collecting and precisely annotating massive LS-SOD data. To advance the research and evaluation in this area, we elaborately collect the first new real LS-SOD dataset, termed Low light Salient Pedestrian/Vehicle (LSPV) dataset. LSPV comprises 3,100 low-light small pedestrians/vehicles images and covers diverse, challenging cases (e.g., low-light, non-uniform illumination environment, and small objects). Meanwhile, to mitigate the large-scale training data scarcity and avoid laborious manual labeling, we proposed a Scale-Illumination (SI) data augmentation method to easily create infinite LS-SOD samples with diverse illumination and salient object scale sizes. Another reason for the under-exploration of LS-SOD is the technical difficulty posed by partially low contrast and limited visual information of small objects in low light, which hinders existing SOD methods from accurately locating and segmenting salient objects. To this end, we propose a baseline LS-SOD network named Illumination and Edge-Driven Network (IEDNet), which explicitly learns illumination and edge features to guide saliency detection. Furthermore, a practical Crop-Fusion (CF) post-processing strategy is further proposed to refine the initial saliency maps. Extensive experiments show that our SI and CF strategies significantly improve current SOD models' performance on the LS-SOD dataset. Moreover, our method achieves state-of-the-art performance on both the real LS-SOD and DUTS-TE datasets.

Lighting Task analysis Object detection Training Image edge detection Feature extraction Data augmentation Low-light sampling generation salient object detection small object

SCI

英语

其他

Natural Science Foundation of China["62376201","U1803262","61602349","61440016"]

Computer Science

Computer Science, Artificial Intelligence

WOS:001303427100001

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

Web of Science

1.Wuhan Univ Sci & Technol, Sch Comp Sci & Technol, Wuhan 430065, Peoples R China
2.Wuhan Univ, Natl Engn Res Ctr Multimedia Software, Sch Comp Sci, Wuhan 430072, Peoples R China
3.Harbin Inst Technol, Sch Comp Sci & Technol, Harbin 150000, Peoples R China
4.Southern Univ Sci & Technol, Dept Comp Sci & Engn, Guangdong Key Lab Brain Inspired Intelligent Compu, Shenzhen 518055, Peoples R China

Wang, Shiqin,Xu, Xin,Chen, Haoyang,et al. Low-Light Salient Object Detection Meets the Small Size[J]. IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE,2024.

Wang, Shiqin,Xu, Xin,Chen, Haoyang,Jiang, Kui,Wang, Zheng,&Tang, Ke.(2024).Low-Light Salient Object Detection Meets the Small Size.IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE.

Wang, Shiqin,et al."Low-Light Salient Object Detection Meets the Small Size".IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE (2024).