Reproduce “EMLight: Lighting Estimation via Spherical Distribution Approximation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fangneng Zhan, Changgong Zhang, Yingchen Yu, Yuan Chang, Shijian Lu, Feiying Ma, Xuansong Xie
EMLight, a framework from Nanyang Technological University and Alibaba DAMO Academy, estimates illumination from a single image by decomposing it into a spherical light distribution, intensity, and ambient term, then uses a neural projector with spherical convolution to synthesize panoramic HDR maps. The framework consistently achieved superior performance on the Laval Indoor HDR Dataset, demonstrating the lowest RMSE (0.062) and angular error (6.43 degrees) for diffuse materials, and improving perceived realism in user studies compared to existing methods.
Key insights
- Decomposing complex illumination into spatially distinct components (distribution, intensity, ambient) simplifies the estimation task and allows for more targeted regression.
- The Spherical Mover's Loss (SML) effectively captures the global geometry and spatial relationships of spherical light distributions, leading to more accurate light source localization than traditional L2 or cross-entropy losses.
- Integrating spherical convolutions into the neural projector addresses panoramic image distortions, enabling the generation of high-fidelity illumination maps that retain crucial frequency information for realistic rendering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.