Representation Learning for Object Detection from Unlabeled Point Cloud Sequences

Abstract

Although unlabeled 3D data is easy to collect, state-of-the-art machine learning techniques for 3D object detection still rely on difficult-to-obtain manual annotations. To reduce dependence on the expensive and error-prone process of manual labeling, we propose a technique for representation learning from unlabeled LiDAR point cloud sequences. Our key insight is that moving objects can be reliably detected from point cloud sequences without the need for human-labeled 3D bounding boxes. In a single LiDAR frame extracted from a sequence, the set of moving objects provides sufficient supervision for single-frame object detection. By designing appropriate pretext tasks, we learn point cloud features that generalize to both moving and static unseen objects. We apply these features to object detection, achieving strong performance on self-supervised representation learning and unsupervised object detection tasks.