The paper "The Unreasonable Effectiveness of Deep Features as a Perceptual Metric" explores the effectiveness of deep features in capturing human perceptual similarity between images. Traditional metrics like L2/PSNR, SSIM, and FSIM often fail to align with human judgments, while deep networks, regardless of architecture or supervision type, show strong agreement with human perception. The authors introduce a large-scale dataset of human perceptual similarity judgments, evaluating deep features across various architectures and tasks. They find that deep features outperform traditional metrics significantly, even when trained on unsupervised data. The results suggest that perceptual similarity is an emergent property of deep visual representations, regardless of the training signal. The study also demonstrates that deep features can be calibrated to better match human judgments, indicating that deep learning models can effectively capture human visual perception. The paper highlights the importance of network architecture and training signals in achieving high perceptual similarity, and shows that even simple unsupervised networks can perform well. The findings support the idea that deep features are a powerful tool for perceptual assessment, outperforming traditional metrics in capturing human-like visual similarity.The paper "The Unreasonable Effectiveness of Deep Features as a Perceptual Metric" explores the effectiveness of deep features in capturing human perceptual similarity between images. Traditional metrics like L2/PSNR, SSIM, and FSIM often fail to align with human judgments, while deep networks, regardless of architecture or supervision type, show strong agreement with human perception. The authors introduce a large-scale dataset of human perceptual similarity judgments, evaluating deep features across various architectures and tasks. They find that deep features outperform traditional metrics significantly, even when trained on unsupervised data. The results suggest that perceptual similarity is an emergent property of deep visual representations, regardless of the training signal. The study also demonstrates that deep features can be calibrated to better match human judgments, indicating that deep learning models can effectively capture human visual perception. The paper highlights the importance of network architecture and training signals in achieving high perceptual similarity, and shows that even simple unsupervised networks can perform well. The findings support the idea that deep features are a powerful tool for perceptual assessment, outperforming traditional metrics in capturing human-like visual similarity.