One Point, One Object: Simultaneous 3D Object Detection and 6-DOF Pose Estimation
Hongsen Liu, Xue Zhi Tan, Yiyi Yin, Bin Wang · 2024
We propose an efficient single shot method for simultaneous 3D object detection and 6-DOF pose estimation in pure 3D point clouds scenes based on a consensus that one point only belongs to one object, i.e., each point has the potential power to predict the 6-DOF pose of its corresponding object. Unlike various point clouds based methods for the similar task that converts the point clouds into regular 3D voxel grids to overcome its irregular structure and do voxel-wise prediction, or to segment the set of the target point clouds in advance and predict on the given segmentation, ours is concise enough to solve the point-wise prediction in 3D point clouds without any structure conversion and stepwise processing. The key component of our method is a multi-task segmentation and prediction network, which can simultaneous predicts: (1) the point-wise semantic segmentation for filtering out background points and reducing search space, (2) the 3D locations of the object 3D bounding box vertices for estimating 6-DOF pose transformation and (3) the confidence for evaluating the accuracy of 3D bounding box prediction.