IDT: Multi-View Intrinsic Decomposition with a Physically Grounded Transformer

This paper introduces IDT, a novel feed-forward transformer-based framework for multi-view intrinsic image decomposition. It addresses the challenge of view inconsistency in existing methods by jointly reasoning over multiple input images. The use of a physically grounded image formation model, decomposing images into diffuse reflectance, diffuse shading, and specular shading, is a key contribution, enabling interpretable and controllable decomposition. The focus on multi-view consistency and the structured factorization of light transport are significant advancements in the field.