Appendix for Exploring Forensic Dental Identification with Deep Learning

We apply the domain-specific augmentations of (i) random tooth reduction, (ii) random artifact addition, (iii) random rigid patch transform, as well as (iv) random contrast shifting and Gaussian noise for instance discriminative learning. The DSA is enabled by the anatomical awareness, and its parameters are set by working with dentists to best follow the possible clinical cases. In specific, for random tooth reduction, one tooth area is set to the background intensity according to the tooth mask from semantic segmentation. The background intensity is determined by the average intensity of the non-teeth area. For random artifact addition, two types of artifacts are included: braces and dental filling.