3D Object Detection with Track-Based Auto-Labelling Using Very Sparsely Labelled Data

In the context of LiDAR-based 3D object detection, we consider the problem of generating high-quality pseudo-labels from very sparsely labelled data. We focus on track-based auto-labelling, which is a class of state-of-the-art pseudo-labelling methods that exploits the sequential nature of point cloud collection, but typically expects training data to be densely labelled. In this work, we analyze different ways to adapt a particular track-based auto-labelling approach to sparsely labelled sequential data from the Waymo Open Dataset, valuing balanced performance on stationary and dynamic vehicles. We thus propose methods that achieve high performance on both of these categories, with as few as one labelled frame per sequence.