imagenet-ipc1-nearest-neighbor-denser-than-wit-1024

IN premise — summaries/2026/08/24/koepke-2026-back-into-cave-sR-references.md

Created 2026-08-24T17:11:01+00:00

ImageNet validation set at 1 image per class (ipc=1) yields k=1 nearest-neighbor cosine similarity of 0.823 for DINOv2-base, exceeding WIT-1024's 0.799, confirming ImageNet operates in a similarly dense retrieval regime to WIT-1M.

Summary

When you look at how tightly images cluster in embedding space, a standard ImageNet benchmark shows nearly the same density as a large web-scale image-text dataset, meaning the retrieval and similarity patterns you'd expect from massive real-world collections are already present in smaller curated benchmarks. This matters because results and assumptions developed for web-scale retrieval transfer to ImageNet-style tasks without needing separate treatment.