deep hybrid real and synthetic training for intrinsic decomposition

Department: Computer Science & Engineering
Research Institute Affiliation: Center for Visual Computing
Faculty Advisor(s): Ravi Ramamoorthi

Primary Student
Name: Sai Bi
Phone: 858-352-8857
Grad Year: 2020

Intrinsic image decomposition is the process of separating the reflectance and shading layers of an image, which is a challenging and underdetermined problem. In this paper, we propose to systematically address this problem using a deep convolutional neural network (CNN). Although deep learning has been recently used to handle this application, the current state-of-the-art methods train the network only on synthetic images as obtaining ground truth reflectance and shading for real images is difficult. Therefore, these methods fail to produce reasonable results on real images. We overcome this limitation by proposing a novel hybrid approach to train our network on both synthetic and real images. Specifically, in addition to directly supervising the network using synthetic images, we train the network by enforcing it to produce the same reflectance for a pair of images of the same real-world scene with different illuminations. Furthermore, we improve the results by incorporating a bilateral layer into our system during both training and test stages. Experimental results show that our approach produces better results than the state-of-the-art methods on various synthetic and real datasets both visually and numerically. Additionally, we extend our method to produce temporally coherent intrinsic decomposition on videos.

Industry Application Area(s)
Software, Analytics | Computer Graphics, Machine Learing

« Back to Posters or Search Results