Zheng Wu ; Tang Weiliang ; Chen Sijin ; Jiang Li ; Fu Chi-Wing

Abstract

Existing single-stage detectors for locating objects in point clouds oftentreat object localization and category classification as separate tasks, so thelocalization accuracy and classification confidence may not well align. Toaddress this issue, we present a new single-stage detector named the ConfidentIoU-Aware Single-Stage object Detector (CIA-SSD). First, we design thelightweight Spatial-Semantic Feature Aggregation module to adaptively fusehigh-level abstract semantic features and low-level spatial features foraccurate predictions of bounding boxes and classification confidence. Also, thepredicted confidence is further rectified with our designed IoU-awareconfidence rectification module to make the confidence more consistent with thelocalization accuracy. Based on the rectified confidence, we further formulatethe Distance-variant IoU-weighted NMS to obtain smoother regressions and avoidredundant predictions. We experiment CIA-SSD on 3D car detection in the KITTItest set and show that it attains top performance in terms of the officialranking metric (moderate AP 80.28%) and above 32 FPS inference speed,outperforming all prior single-stage detectors. The code is available athttps://github.com/Vegeta2020/CIA-SSD.