2021/03/24 by Jonathan T. Barron, Barron, Jonathan T., Ben Mildenhall +9 · 168 citations
Medicine · Computer Science · #Advanced Neuroimaging Techniques and Applications #Advanced Neural Network Applications #Medical Image Segmentation Techniques
paper · pdf · doi:10.48550/arxiv.2103.13415
The rendering procedure used by neural radiance fields (NeRF) samples a scene\nwith a single ray per pixel and may therefore produce renderings that are\nexcessively blurred or aliased when training or testing images observe scene\ncontent at different resolutions. The straightforward solution of supersampling\nby rendering with multiple rays per pixel is impractical for NeRF, because\nrendering each ray requires querying a multilayer perceptron hundreds of times.\nOur solution, which we call "mip-NeRF" (a la "mipmap"), extends NeRF to\nrepresent the scene at a continuously-valued scale. By efficiently rendering\nanti-aliased conical frustums instead of rays, mip-NeRF reduces objectionable\naliasing artifacts and significantly improves NeRF's ability to represent fine\ndetails, while also being 7% faster than NeRF and half the size. Compared to\nNeRF, mip-NeRF reduces average error rates by 17% on the dataset presented with\nNeRF and by 60% on a challenging multiscale variant of that dataset that we\npresent. Mip-NeRF is also able to match the accuracy of a brute-force\nsupersampled NeRF on our multiscale dataset while being 22x faster.\n