The Evolution of Image Based 3D Reconstruction

This episode of Computer Vision Decoded features a presentation by Jared Heinly on the evolution of 3D reconstruction. Jared discusses the history of 3D reconstruction, starting from early methods like stereoscopes and metrophotography to modern techniques involving computer vision, machine learning, and AI. He explains active and passive sensing, focusing on image-based passive sensing, and details the progression from manual stereo plotters to automated systems using bundle adjustment and structure from motion. The talk covers advanced concepts like monocular depth estimation, neural radiance fields (NeRF), and Gaussian splatting, highlighting how these methods leverage machine learning to create detailed 3D representations from images. Jared also touches on the use of crowdsourced photo collections and various 3D data representations, emphasizing the ongoing evolution and diverse applications of 3D reconstruction in computer vision.