Knowledge Distillation for Predicting Varying Environment Maps from Single Images

SHANG-LIN LU, Ting-I Hsieh, Hwann-Tzong Chen

Keywords: Environment Maps, Intrinsic Images

Abstract:

We present a learning-based method for estimating view-dependent environmental lighting from a single image. Our method learns to distill knowledge from a differentiable geometry and texture decomposition framework. The goal is to directly predict the environment map from an input image using a neural network and thus bypass the need for solving iterative optimization.
We propose a new physically-based strategy that decouples the illumination color and distribution of a local light probe given by a sampled pixel on the input image. The experiments show that our proposed method can train a neural network to efficiently derive the environment map of comparable or even higher quality from a single image in under a second, a significant improvement over the time-consuming optimization-based alternatives that require a few minutes to obtain the results.

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

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: 984