FPN++: A Simple Baseline for Pedestrian Detection
Junhao Hu, Lei Jin, Shenghuo Gao · 2019
Our observation shows that pedestrians' heights greatly affect pedestrian detection performance, and for small pedestrians, their context information is useful for localizing and recognizing these pedestrians. Based on our observation, a FPN++ framework, which is an extension of Feature Pyramid Network (FPN) is proposed. It improves the FPN from the following aspects: i) we modify the backbone of FPN by reducing the stride of convolution from 2 to 1 in FPN from earlier layers, which allows the network to detect smaller pedestrians with more semantically meaningful features extracted from deeper layers, then we replace the convolution with dilated convolution to increase the local receptive fields and facilitate the detection on pedestrians of all scales; ii) a context-aware detection module is introduced in the predictor head of FPN to leverage context information for detection. Extensive experiments on the CityPersons and Caltech pedestrian datasets show that our FPN++ achieves state-of-the-art performance and significantly improves the performance for small pedestrians. Our solution can be readily extended to other detection tasks, and experiments on the VOC2007 benchmark also validate the effectiveness of our solution.