Low-Light Salient Object Detection Meets the Small Size

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
L
ow light
S
alient
P
edestrian/
V
ehicle (
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.