A Comprehensive Real-World Photometric Stereo Dataset for Unsupervised Anomaly Detection

Unsupervised anomaly detection is a technology that trains a model to detect anomalies using only normal data. One of the main challenges of anomaly detection is detecting defects that are not visible or obscured by lighting conditions. However, most existing methods use a single image as an input to detect anomalies in the target object, thus it is difficult to properly cope with such a challenge. In this paper, we explore an anomaly detection model using multiple images taken under different lighting conditions. For this purpose, we propose a new anomaly detection dataset based on the photometric stereo (PS) setup, and demonstrate its characteristics and advantages. In addition, we apply existing anomaly detection methods to our new dataset to verify their performance and present benchmark results. Furthermore, we present a technique for fusing multi-image and 3D surface information. Anomaly detection based on our new setup using multiple images achieves dramatic improvement of
$+10.15\%$
,
$+11.68\%$
and
$+4.38\%$
over baseline setup using a single image for Skip-GANomaly, MKDAD and DifferNet models, respectively. Our PS-based anomaly detection dataset and analysis reports will be publicly available for future researches.