Evidence · 3 · Measure & Manufacture

Slide 3 under the hood

Under the hood, matching means comparing features

Every matcher, from SIFT to cost volumes to transformers, scores similarity between local descriptors and prefers the best. The slide used no correspondence training at all: self-supervised DINOv2 features and plain cosine similarity. The sparse matches on the slide are read straight out of that similarity field, and eyes, nose and chin land correctly. What supervised matchers add is the rest: sharper features, and the ability to turn a noisy similarity map into a coherent field.

Interactive · try it hereOpen standalone ↗

Matching, Live

Hover or tap anywhere on the left image. The heat map is the cosine similarity of that patch's DINOv2 feature to every patch of the right image; the ring is the best match. Nothing here was trained for correspondence.

Image pair

View

sharpness0.08

move over the left image

Features: DINOv2 ViT-B/14 patch tokens (32×32 grid at 448 px), reduced to 64 dimensions. The cat, horse and dog pairs are PF-PASCAL test pairs with their annotated keypoints; the KITTI pair is two consecutive driving frames, where the dense field becomes optical flow. Similarity is a plain dot product. The dense field applies a soft-argmax with the sharpness above and a median filter, exactly the recipe behind the slide's correspondence field. Try the left cat's eye: the right cat's eye lights up; try the other eye and the match lands on the wrong eye, which is the honest part: symmetric parts are where zero-shot features stop and supervised training begins.