Weakly Supervised Point Clouds Transformer for 3D Object Detection

The annotation of 3D datasets is required for semantic-segmentation and\nobject detection in scene understanding. In this paper we present a framework\nfor the weakly supervision of a point clouds transformer that is used for 3D\nobject detection. The aim is to decrease the required amount of supervision\nneeded for training, as a result of the high cost of annotating a 3D datasets.\nWe propose an Unsupervised Voting Proposal Module, which learns randomly preset\nanchor points and uses voting network to select prepared anchor points of high\nquality. Then it distills information into student and teacher network. In\nterms of student network, we apply ResNet network to efficiently extract local\ncharacteristics. However, it also can lose much global information. To provide\nthe input which incorporates the global and local information as the input of\nstudent networks, we adopt the self-attention mechanism of transformer to\nextract global features, and the ResNet layers to extract region proposals. The\nteacher network supervises the classification and regression of the student\nnetwork using the pre-trained model on ImageNet. On the challenging KITTI\ndatasets, the experimental results have achieved the highest level of average\nprecision compared with the most recent weakly supervised 3D object detectors.\n