作者: Evan Shelhamer , Jonathan T. Barron , Trevor Darrell
关键词:
摘要: Intrinsic image decomposition factorizes an observed into its physical causes. This is most commonly framed as a reflectance and shading, although recent progress has made full decompositions shape, illumination, reflectance, shading possible. However, existing factorization approaches require depth sensing to initialize the optimization of scene intrinsics. Rather than relying on sensors, we show that estimated purely from monocular appearance can provide sufficient cues for intrinsic analysis. Our pipeline regresses by fully convolutional network then jointly optimizes recover input image. combination yields uniting feature learning through deep regression with modeling statistical priors random field regularization. work demonstrates first scenes single color alone.