Local Implicit Ray Function for Generalizable Radiance Field Representation

Figure 1 . We propose LIRF to reconstruct radiance fields of unseen scenes for novel view synthesis. Given that current generalizable NeRF-like methods cast an infinitesimal ray to render a pixel at different scales, it causes excessive blurring and aliasing. Our method instead reasons about 3D conical frustums defined by the neighbor rays through the neighbor pixels (as shown in (a)). Our LIRF outputs the feature of any sample within the conical frustum in a continuous manner (as shown in (b)), which supports NeRF reconstruction at arbitrary scales. Compared with the previous method, our method can be generalized to represent the same unseen scene at multiple levels of details (as shown in (c)). Specifically, given a set of input views at a consistent image scale ×1, LIRF enables our method to both preserve sharp details in close-up shots (anti-blurring as shown in ×2 and ×4 results) and correctly render the zoomed-out images (anti-aliasing as shown in ×0.5 results).