Kumar Abhinav ; Brazil Garrick ; Liu Xiaoming

Abstract

Modern 3D object detectors have immensely benefited from the end-to-endlearning idea. However, most of them use a post-processing algorithm calledNon-Maximal Suppression (NMS) only during inference. While there were attemptsto include NMS in the training pipeline for tasks such as 2D object detection,they have been less widely adopted due to a non-mathematical expression of theNMS. In this paper, we present and integrate GrooMeD-NMS -- a novel GroupedMathematically Differentiable NMS for monocular 3D object detection, such thatthe network is trained end-to-end with a loss on the boxes after NMS. We firstformulate NMS as a matrix operation and then group and mask the boxes in anunsupervised manner to obtain a simple closed-form expression of the NMS.GrooMeD-NMS addresses the mismatch between training and inference pipelinesand, therefore, forces the network to select the best 3D box in adifferentiable manner. As a result, GrooMeD-NMS achieves state-of-the-artmonocular 3D object detection results on the KITTI benchmark dataset performingcomparably to monocular video-based methods. Code and models athttps://github.com/abhi1kumar/groomed_nms