Diskriminativna metoda za detekcijo 3D anomalij

Recent methods for surface anomaly detection are based on feature extraction using pretrained networks. On RGB anomaly detection datasets these methods achieve excellent results, but the results for 3D anomaly detection are worse due to the lack of pretrained networks suitable for this domain. There is also a lack of industrial depth image datasets that allow learning of networks that could be used in these methods. Discriminative anomaly detection methods do not require pretrained networks and learn with simulated anomalies. The process of simulating anomalies appropriate to the domain of industrial depth data is non-trivial and is necessary for training discriminative methods. We propose a novel 3D anomaly simulation process that is suitable for learning discriminative methods. We demonstrate the effectiveness of the process using DRÆM-3D, a strong discriminative method for 3D anomaly detection. The proposed approach achieves excellent results on the MVTec3D anomaly detection database, where DRÆM-3D outperforms all previous state-of-the-art methods on both the 3D anomaly detection problem and the 3D+RGB anomaly detection problem.