Dual Subnet Label Assignment for End-to-End Fully Convolutional Pedestrian Detection in Crowd Scenes

The fully convolutional detector uses a one-to-one (O2O) label assignment strategy removing NMS post-processing operations and realizes end-to-end detection. However, limited positive samples lead to slow convergence of the full convolutional end-to-end detector in crowded scenes. In this paper, we propose a Dual Subnet Label Assignment network (DSLA), which accelerates the model convergence to improve the detection performance and achieves end-to-end detection pipeline. First, we present Soft Label Assignment (SLA), which utilizes soft anchors with positive and negative sample semantics for model learning and accelerates model convergence. Meanwhile, SLA assigns labels to occluded targets in crowded scenes. We combine the SLA branch and the O2O branch for co-training, and achieve end-to-end detection through the O2O branch. Finally, we propose Feature Shuffle to solve the problem of lack of information interaction between different feature layers and highlight the features of the occluded part in a crowded scenes. Experiments demonstrate that DSLA outperforms OneNet in terms of model convergence speed and detection performance. Especially, on the CrowdHuman dataset, our method outperforms SOTA, achieving 91.7 AP, 47.2 MR⁻², 80.3 JI, and 98.5 Recall.