HDR-NeRF

Summary

We current Excessive Dynamic Vary Neural Radiance Fields (HDR-NeRF) to recuperate an HDR radiance subject from a set of low dynamic vary (LDR) views with completely different exposures. Utilizing the HDR-NeRF, we’re capable of generate each novel HDR views and novel LDR views below completely different exposures. The important thing to our methodology is to mannequin the bodily imaging course of, which dictates that the radiance of a scene level transforms to a pixel worth within the LDR picture with two implicit features: a radiance subject and a tone mapper. The radiance subject encodes the scene radiance (values fluctuate from 0 to +infty), which outputs the density and radiance of a ray by giving corresponding ray origin and ray route. The tone mapper fashions the mapping course of {that a} ray hitting on the digicam sensor turns into a pixel worth. The colour of the ray is predicted by feeding the radiance and the corresponding publicity time into the tone mapper. We use the traditional quantity rendering method to undertaking the output radiance, colours, and densities into HDR and LDR photos, whereas solely the enter LDR photos are used because the supervision. We gather a brand new forward-facing HDR dataset to judge the proposed methodology. Experimental outcomes on artificial and real-world scenes validate that our methodology cannot solely precisely management the exposures of synthesized views but additionally render views with a excessive dynamic vary.