UNION, a novel unsupervised 3D object detection method, leverages the joint strengths of LiDAR and camera data to achieve state-of-the-art performance by effectively distinguishing between static foreground and background objects, eliminating the need for computationally expensive self-training.
Inaccurate pseudo bounding boxes hinder unsupervised 3D object detection, but a novel uncertainty-aware framework, UA3D, mitigates this by estimating and regularizing uncertainty at the coordinate level, leading to substantial performance improvements.
This paper introduces a Commonsense Prototype-based Detector (CPD) that can perform accurate unsupervised 3D object detection without human annotations. CPD constructs high-quality commonsense prototypes to refine pseudo-labels and guide the network convergence, significantly outperforming state-of-the-art unsupervised 3D detectors.