A New Perspective of Weakly Supervised 3D Instance Segmentation via Bounding Boxes

Abstract

Existing fully supervised method 3D point cloud segmentation methods heavily rely on carefully annotated point labels. In this work, we look at weakly-supervised 3D instance segmentation using bounding boxes supervision. Bounding boxes are much easier to annotate than dense point-wise labels. Moreover, they demonstrated high potential in addressing instance-level segmentation compared to other types of weak annotations. However, existing bounding-box supervised techniques have struggled to keep pace with the development of fully-supervised methods. To tackle this issue, we propose a simple-yet-effective approach to directly leverage the network architecture of fully-supervised methods for such weak supervision scenarios. We found that accurate instance labels for each point can be generated with the given bounding boxes by leveraging 3D geometric prior. Such a process is efficient and does not require any additional training or fine-tuning. The generated point-wise labels can be fed to any advanced fully-supervised model without re-designing specific networks for bounding-box supervision. In this fashion, our designed approach achieves on par performance of fully supervised methods in terms of AP, AP50 and AP25. Remarkably, we outperformed the state-of-the-art bounding-box supervised method by 21%. Compared with existing methods, our method is extremely simple and only involves two small heuristics in the data preprocessing step. In addition, our method is proven to be robust against noisy bounding box scenario through experiments.