This paper presents a new method for 3D object localization using only 2D labels and the physical laws of motion. No 3D labels or expensive hardware is needed. The model is trained with easy-to-annotate 2D labels, and physical knowledge of the object’s motion, allowing it to infer the latent third dimension. The method is tested on both synthetic and real-world datasets, achieving a mean distance error of just 6 cm in real data experiments. This method has potential for applications where collecting 3D data for training is not feasible, such as low budget sports.
Publication date: 27 Oct 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2310.17462
