2 Dimensional Label

Two-dimensional (2D) labels, readily obtained from images, are increasingly used to train models for tasks traditionally requiring expensive 3D annotations. Current research focuses on leveraging 2D labels through weakly supervised learning techniques, often incorporating self-supervised components, pseudo-labeling strategies, and multi-view consistency constraints to improve 3D model accuracy. This approach is significant because it reduces the reliance on costly 3D data annotation, enabling the training of more sophisticated models for applications like 3D object detection, scene understanding, and activity recognition in areas such as autonomous driving and robotics.