Progressively Optimized Local Radiance Fields for Robust View Synthesis

Input: casually captured long video Output: jointly estimated camera poses and local radiance fields LocalRF (ours): high-quality novel view synthesis BARF [17]: the estimated poses often fall into local minima for long sequences Mip-NeRF360 [4]: the spatial resolution is often limited throughout the video Figure 1 . High-quality novel view synthesis from a long casually captured video. We jointly optimize camera poses and a scene representation using a progressive scheme that dynamically allocates local radiance fields (blue boxes). Our method robustly handles casual hand-held captures, scales to processing arbitrarily long videos with limited memory usage, and maintains high resolution throughout the entire video.