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

Abstract

We present Self-Ensembling Single-Stage object Detector (SE-SSD) for accurateand efficient 3D object detection in outdoor point clouds. Our key focus is onexploiting both soft and hard targets with our formulated constraints tojointly optimize the model, without introducing extra computation in theinference. Specifically, SE-SSD contains a pair of teacher and student SSDs, inwhich we design an effective IoU-based matching strategy to filter soft targetsfrom the teacher and formulate a consistency loss to align student predictionswith them. Also, to maximize the distilled knowledge for ensembling theteacher, we design a new augmentation scheme to produce shape-aware augmentedsamples to train the student, aiming to encourage it to infer complete objectshapes. Lastly, to better exploit hard targets, we design an ODIoU loss tosupervise the student with constraints on the predicted box centers andorientations. Our SE-SSD attains top performance compared with all priorpublished works. Also, it attains top precisions for car detection in the KITTIbenchmark (ranked 1st and 2nd on the BEV and 3D leaderboards, respectively)with an ultra-high inference speed. The code is available athttps://github.com/Vegeta2020/SE-SSD.