Stereo VoVNet-CNN for 3D object detection

Kaiqi Su, Weiqing Yan, Xin Wei, Meiqi Gu

Abstract: 3D object detection is a key issue and research in autonomous vehicle and computer vision. 3D detection methods based on stereoscopic images estimate 3D boxes and regress the object pose by exploiting the sparse and dense, semantic and geometry information of stereoscopic images. In this paper, we propose stereo VoVNet-CNN for 3D object detection. In this network, we adopt VoVNet as the backbone to extract feature, and Stereo Region Proposal Network(RPN) to refine 2D box, moreover, align and extract 3D box by stereo regression with 3D constraint of stereo images. The VoVNet module considers local and global feature information by integrating the contextual information from different receptive fields and inputting initial feature into the final output feature map. Therefore, the proposed stereo VoVNet can provide high accuracy for feature extraction. We evaluate the proposed network on KITTI dataset by comparing it with other state-of-the-art methods. Experimental results demonstrate well the effectiveness of the proposed network.