The first WIDER Face and Pedestrian Challenge was held at the European Conference on Computer Vision (ECCV2018), hosted by SenseTime Technology, Chinese University of Hong Kong, Amazon, Nanyang Technological University and the University of Sydney. During the competition, it attracted more than 400 participants from all over the world with final results from 73 teams.
In the Pedestrian Detection task, the team of Visual Information Processing and Learning (VIPL) from the Institute of Computing Technology, Chinese Academy of Sciences won first place (1/35). Based on the architecture of Faster RCNN, we adopted ResNet with Feature Pyramid Networks (FPN) to extract and fused multi-level semantic features. Better regression results are obtained by adding a cascade strategy (Cascade R-CNN) to the detection module. RoI-Align is adopted instead of RoI-Pooling in the original Faster RCNN for detecting small-scale pedestrians and channel-wise attention is used to deal with occlusion. The overall framework is shown in Figure 1. VIPL team ranked first on the leaderboard, exceeding the second place by five percentage points (Figure 2). At the challenge workshop, the VIPL team received the award certificate issued by the organizer (Figure 3) and gave a 10-minute oral report (Figure 4).
Figure 1: Overall Architecture
Figure 2: Results
Figure 3: Award Certificate，team members：Hongkai Zhang, Bingpeng Ma, Hong Chang, Shiguang Shan, Xilin Chen.
Figure 4: Oral report by Prof. Shiguang Shan
Download: