Neural Architecture Search for Efficient Uncalibrated Deep Photometric Stereo

We present an automated machine learning approachfor uncalibrated photometric stereo (PS). Our work aimsat discovering lightweight and computationally efficient PSneural networks with excellent surface normal accuracy.Unlike previous uncalibrated deep PS networks, which arehandcrafted and carefully tuned, we leverage differentiableneural architecture search (NAS) strategy to find uncali-brated PS architecture automatically. We begin by defininga discrete search space for a light calibration network anda normal estimation network, respectively. We then performa continuous relaxation of this search space and presenta gradient-based optimization strategy to find an efficientlight calibration and normal estimation network. Directlyapplying the NAS methodology to uncalibrated PS is notstraightforward as certain task-specific constraints must besatisfied, which we impose explicitly. Moreover, we searchfor and train the two networks separately to account forthe Generalized Bas-Relief (GBR) ambiguity. Extensive ex-periments on the DiLiGenT dataset show that the automat-ically searched neural architectures performance comparesfavorably with the state-of-the-art uncalibrated PS methodswhile having a lower memory footprint.

• Machine Perception