PreFIQs: Face Image Quality Is What Survives Pruning
Researchers have introduced PreFIQs, a novel, unsupervised framework for assessing face image quality. This method leverages the Pruning Identified Exemplar (PIE) hypothesis, suggesting that low-utility images have embeddings that are more sensitive to model pruning. PreFIQs quantifies image utility by measuring the distance between embeddings from a full model and its pruned version, offering a training-free approach that achieves competitive or state-of-the-art results on multiple benchmarks. AI
IMPACT Introduces a novel, training-free method for evaluating face image utility, potentially improving downstream face recognition systems.