Deep intrinsic image decomposition under colored AC light sources

Intrinsic image decomposition assumes that the observed color image can be decomposed into reflectance and illumination. It is beneficial for understanding the physical world, but a severely ill-posed problem. Several sequence-based deep learning methods only exploit spatial prior to illumination, while the proposed method introduced temporal prior of illumination. They also assume gray illumination which may cause color distortion in the reflectance image. This paper proposes a deep intrinsic image decomposition method using a high-speed camera under colored AC light sources. A high-speed camera can capture the sinusoidal variations in scene brightness, which was used to extract the temporal correlation among high-speed video frames. With these powerful cues, the proposed method jointly performs intrinsic image decomposition and color constancy. To the best of our knowledge, this is the first study that exploits AC light properties for intrinsic image decomposition. We evaluate the color constancy and intrinsic image decomposition quality to validate the model estimation accuracy. The experimental results show that the proposed deep network can accurately estimate both illumination color and intrinsic images, and the two factors are mutually supportive each other for learning.