The content introduces IFFNeRF, a method for estimating the camera pose of an image using Neural Radiance Fields (NeRF) in real-time. It eliminates the need for an initial pose guess and improves accuracy by 80.1% in angular error and 67.3% in translation error compared to iNeRF. The process involves Metropolis-Hasting algorithm for surface point sampling, ray casting from sampled points, attention mechanism for matching rays with image pixels, and Least Squares optimization for pose estimation. Evaluation on synthetic and real datasets shows superior performance over iNeRF, especially without constrained initialization.
翻譯成其他語言
從原文內容
arxiv.org
深入探究