Zhang et al. (2018) — ARXIV-1801.03924
Key takeaways
- A large human perceptual-similarity dataset is collected first, so the question “which metric agrees with people?” becomes answerable rather than argued.
- Against it, distances computed in the feature space of a pretrained network outperform every classical metric — L2, SSIM, FSIM — by a wide margin. The metric is LPIPS.
- The striking part is the insensitivity: the result holds across architectures (SqueezeNet, AlexNet, VGG) and across supervised, self-supervised and even unsupervised pretraining. Perceptual structure appears to be a general property of learned visual representations rather than a property of one training signal.
- Calibrating a small linear layer on top of the features improves agreement further, which is how LPIPS is normally used.
Standing in the anthology
Nothing in the record cites this note, and the anthology is about training rather than evaluation, so it carries no practice. It is filed as a real paper rather than retired because the corpus’s generative-vision half (LIT-070, LIT-073, LIT-118) reports FID and perceptual distances throughout, and this is the paper those numbers rest on.
Read — NOTE-183. Still no practice, and the reading says why more precisely than “the anthology is about training”: the paper’s most transferable claim is a negative one. The effect is not VGG’s, not ImageNet’s and not supervision’s — it survives losing all three — but it is also not architecture’s alone, because a randomly initialized network does not produce it. What is left is “train on something”, which is not an instruction anybody needs.