SPCS: a spatial pyramid convolutional shuffle module for YOLO to detect occluded object

Abstract In crowded scenes, one of the most important issues is that heavily overlapped objects are hardly distinguished from each other since their pixels shared and visible occluded objects, which used to represent features, limited. this paper, a spatial pyramid convolutional shuffle (SPCS) module proposed extract refined information limited generate distinguishable representations for objects. We adopt four kernels with different sizes dilation rates at location in features adjacently recombine fused outputs spatially using pixel module. way, instance predictions corresponding can be produced feature. addition, multiple operations kernel same regions, helpful pixels. Extensive experimental results demonstrate SPCS effectively boost performance human detection. YOLO detector achieves 94.11% AP, 41.75% MR, 97.75% Recall on CrowdHuman, 93.04% 98.45% WiderPerson, best compared previous state-of-the-art models.