LLA: Loss-aware label assignment for dense pedestrian detection

• DOI: 10.1016/j.neucom.2021.07.094
• Publication Date: 2021-08-06T15:34:03Z
• Authors: Zheng Ge; Jianfeng Wang; Xin Huang; Songtao Liu; Osamu Yoshie
• Keywords: Pedestrian detection; CLs upper limits; Ground truth

Label assignment has been widely studied in general object detection because of its great impact on detectors' performance. In the field dense pedestrian detection, human bodies are often heavily entangled, making label more important. However, none existing method focuses crowd scenarios. Motivated by this, we propose Loss-aware Assignment (LLA) to boost performance detectors Concretely, LLA first calculates classification (cls) and regression (reg) losses between each anchor ground-truth (GT) pair. A joint loss is then defined as weighted summation cls reg assigning indicator. Finally, anchors with top K minimum for a certain GT box assigned positive anchors. Anchors that not any considered negative. simple but effective. Experiments CrowdHuman CityPersons show such strategy can MR 9.53% 5.47% two famous one-stage – RetinaNet FCOS, becoming detector surpasses Faster R-CNN