IMPROVING SEMI-SUPERVISED 3D OBJECT DETECTION

Abstract

3D object detection currently lacks large-scale datasets compared to other tasks such as object classification and 2D object detection. Previous works solve this task by leveraging semi-supervised learning via pseudo labeling and consistency regularisation respectively. Pseudo labeling methods face difficulty in accurately regressing amodal bounding boxes for occluded or incomplete 3D objects as they enforce strong supervision despite over-confident predictions. Existing consistency regularisation-based methods suffer from overwhelming misalignments between the student and teacher proposals in addition to the aforementioned over-confidence issue of the teacher predictions. In this thesis, we tackle these problems by proposing a hybrid filtering mechanism that consists of an objectness-aware hard filtering and a soft filtering for the teacher predictions. The hybrid filtering mechanism can reject potential background proposals produced by the teacher and simultaneously consider the reliability of the remaining proposals. Furthermore, we design a knowledge replication paradigm to iteratively imprint full knowledge from the teacher to the student. We empirically find that such a scheme produces a better student that subsequently benefits teacher learning. We conduct extensive experiments on two benchmark 3D detection datasets and show clear performance improvements over previous state-of-the-art methods.