Triangulation Learning Network: from Monocular to Stereo 3D Object Detection

In this paper, we study the problem of 3D object detection from stereoimages, in which the key challenge is how to effectively utilize stereoinformation. Different from previous methods using pixel-level depth maps, wepropose employing 3D anchors to explicitly construct object-levelcorrespondences between the regions of interest in stereo images, from whichthe deep neural network learns to detect and triangulate the targeted object in3D space. We also introduce a cost-efficient channel reweighting strategy thatenhances representational features and weakens noisy signals to facilitate thelearning process. All of these are flexibly integrated into a solid baselinedetector that uses monocular images. We demonstrate that both the monocularbaseline and the stereo triangulation learning network outperform the priorstate-of-the-arts in 3D object detection and localization on the challengingKITTI dataset.