OpenIllumination: A Multi-Illumination Dataset for Inverse Rendering Evaluation on Real Objects

TL;DR

OpenIllumination tackles the lack of real-object, multi-illumination data for evaluating inverse rendering. It introduces a light-stage–based dataset with 108K images of 64 objects from 72 views under diverse illuminations, plus ground-truth illumination, camera calibration, and masks. The dataset enables quantitative evaluation of inverse rendering and relighting methods, with baseline experiments across single and multi-illumination settings and auxiliary tasks like photometric stereo and novel-view synthesis. The work demonstrates that high-quality real data and precise calibration are essential for robust evaluation and will catalyze progress in real-world material decomposition and relighting.

Abstract

We introduce OpenIllumination, a real-world dataset containing over 108K images of 64 objects with diverse materials, captured under 72 camera views and a large number of different illuminations. For each image in the dataset, we provide accurate camera parameters, illumination ground truth, and foreground segmentation masks. Our dataset enables the quantitative evaluation of most inverse rendering and material decomposition methods for real objects. We examine several state-of-the-art inverse rendering methods on our dataset and compare their performances. The dataset and code can be found on the project page: https://oppo-us-research.github.io/OpenIllumination.

Figures (8)

• Figure 1: Some example images in the proposed dataset. The dataset contains images of various objects with diverse materials, captured under different views and illuminations. The leftmost column visualizes several different illumination patterns, with red and yellow indicating activated and deactivated lights. The name and material for each object are listed in the first and second rows. The materials are selected from the OpenSurfaces bell2013opensurfaces dataset.
• Figure 2: (a) The capturing system contains 48 DSLR cameras (Canon EOS Rebel SL3), 24 high-speed cameras (HR-12000SC), and 142 controllable linear polarized LED. (b) The calibrated DSLR camera poses. (c) The reconstructed light positions.
• Figure 3: The object reconstruction on our dataset from three inverse rendering baselines under single illumination. Objects highlighted by green color are easier tasks in our dataset, while objects in red color are more difficult tasks that involve more complicated materials like metal and clear plastic.
• Figure 4: Relighting results of TensoIR under novel illumination. We show the reconstructed albedo, normal, and PBR results. For each novel illumination, we show the rendering and ground-truth captured images.
• Figure 5: Results of photometric stereo using the OLAT images in our dataset.