Radiance Surfaces: Optimizing Surface Representations with a 5D Radiance Field Loss | RGL

Our meth­od re­con­structs sur­faces with the speed and ro­bust­ness of NeRF-style meth­ods. Left: In con­trast to volume-based meth­ods that min­im­ize 2D im­age losses, as shown in (a), we ad­opt a spa­tio-dir­ec­tion­al ra­di­ance field loss for­mu­la­tion, as shown in (b). At each step, our meth­od con­siders a dis­tri­bu­tion of op­tic­ally in­de­pend­ent sur­faces, in­creas­ing the con­fid­ence of can­did­ates that agree with the ref­er­ence im­agery. Right: A mean­ing­ful sur­face can be ex­trac­ted at any it­er­a­tion dur­ing op­tim­iz­a­tion.

We present a fast and simple tech­nique to con­vert im­ages in­to a ra­di­ance sur­face-based scene rep­res­ent­a­tion. Build­ing on ex­ist­ing ra­di­ance volume re­con­struc­tion al­gorithms, we in­tro­duce a subtle yet im­pact­ful modi­fic­a­tion of the loss func­tion re­quir­ing changes to only a few lines of code: in­stead of in­teg­rat­ing the ra­di­ance field along rays and su­per­vising the res­ult­ing im­ages, we pro­ject the train­ing im­ages in­to the scene to dir­ectly su­per­vise the spa­tio-dir­ec­tion­al ra­di­ance field.

Leave a Comment