richzhang/PerceptualSimilarity

Computes learned perceptual similarity between image patches using pretrained deep networks (AlexNet, VGG, or SqueezeNet) with linearly calibrated intermediate activations, outperforming traditional metrics like SSIM on human judgments. Includes the BAPPS dataset with 95k+ human-annotated triplets for training and evaluation across tasks like super-resolution, deblurring, and color correction. Supports both metric evaluation and backpropagation for use as a perceptual loss in optimization pipelines via PyTorch.

4,185 stars. No commits in the last 6 months.