Continuously Volumetric Rendering with Neural Density-Distance Fields
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: NeRF, Neural Fields, Volumetric Rendering
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: This paper proposes a continuous volumetric rendering and a bisection sampling utilizing the Neural Density-Distance Field (NeDDF) that can synthesize novel views with bouncing transparency during each rendering segment. Since, unlike the density field, the distance field retains the state of the nearby free space, efficient sampling, such as sphere tracing, has been attempted by assuming a solid object. However, distance fields struggle to represent transparency, detailed shapes, and distant landscapes. We derive bounds on transparency in the interval in volume rendering based on NeDDF, which extends distance fields to non-solids. Through realizing the derivation, we invent an efficient bisectional exploratory sampling method that minimizes the maximum of the bound range. For scaling to fit the Eikonal constraints on distance fields, Multi-resolution Hash Encoding, which is excellent for detailed description, is used with frequency separation. We achieve unmasked acquisition of scenes with distant scenery by introducing contract coordinates and scaling the distance field so finite values can describe it. Experiments on synthetic and real data show that the proposed rendering bounds work reasonably.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1522
Loading