Revisiting Face Recognition for Monozygotic Twins: The Celeb Twins Test Set
The paper introduces CTTS, a celebrity-twin benchmark built to test where face matchers still fail on identical twins.
The set covers web-scraped image pairs for 80 monozygotic twin pairs, with metadata on distinguishing skin marks and possible mirror asymmetry. The authors report that current deep CNN face matchers reach above 76% accuracy on same-person versus different-person CTTS pairs. They also find those matchers are not using the skin-mark or asymmetry cues highlighted in earlier twins literature. The paper closes by considering whether generative AI could help create imagined twin images for training data. ArXiv · AI/CL/LG's note
The set covers web-scraped image pairs for 80 monozygotic twin pairs, with metadata on distinguishing skin marks and possible mirror asymmetry. The authors report that current deep CNN face matchers reach above 76% accuracy on same-person versus different-person CTTS pairs. They also find those matchers are not using the skin-mark or asymmetry cues highlighted in earlier twins literature. The paper closes by considering whether generative AI could help create imagined twin images for training data. ArXiv · AI/CL/LG's note
score 3