2 Dimensional Supervision

Two-dimensional (2D) supervision in 3D computer vision research focuses on leveraging readily available 2D data (images, annotations) to train models for 3D tasks, mitigating the high cost and scarcity of 3D data. Current research explores techniques like incorporating temporal information, utilizing neural radiance fields (NeRFs) and their variants (e.g., Nerflets), and employing multi-view consistency constraints to improve 3D estimations from limited 2D input. This approach significantly impacts various applications, including biomedical imaging, autonomous driving, and 3D object detection, by enabling the training of accurate 3D models even with limited or expensive 3D ground truth.