This paper proposes a novel high-resolution multi-view dataset of complex multi-illuminant scenes with precise reflectance and shading ground-truth as well as raw depth and 3D point cloud. Our dataset challenges the intrinsic image methods by providing complex coloured cast shadows, highly textured and colourful surfaces, and specularity. This is the first publicly available multi-view real-photo dataset at such complexity with pixel-wise intrinsic ground-truth. In the effort to help evaluating different intrinsic image methods, we propose a new perception-inspired metric based on the reflectance consistency. We provide the evaluation of three intrinsic image methods using our dataset and metric.
CITATION STYLE
Beigpour, S., Ha, M. L., Kunz, S., Kolb, A., & Blanz, V. (2016). Multi-view multi-illuminant intrinsic dataset. In British Machine Vision Conference 2016, BMVC 2016 (Vol. 2016-September, pp. 10.1-10.13). British Machine Vision Conference, BMVC. https://doi.org/10.5244/C.30.10
Mendeley helps you to discover research relevant for your work.