Deep Perceptual Similarity

Deep perceptual similarity (DPS) aims to develop computational methods that accurately reflect human judgment of image similarity, going beyond simple pixel-by-pixel comparisons. Current research focuses on improving the robustness of DPS metrics, particularly by addressing their sensitivity to image misalignments and adapting them to diverse contexts where the definition of "similarity" can vary. This involves exploring modifications to existing deep neural network architectures, such as LPIPS, and investigating how pre-trained networks can be effectively adapted without sacrificing performance on other tasks. Advances in DPS have significant implications for various computer vision applications, including image retrieval, quality assessment, and generative modeling.