FEVER · FNO + local attention embedder
FEVER image embeddings
Setup
Fingerprinting images to find a way to retrieve them.
Pros
Patent pending cross resolution retrieval engine with massive throughput.
Usage
Deduplicate, natural language search, or cluster huge amounts of photos.
FEVER embeds images into a shared image/text space. Its trick: the embedding barely moves as the picture
shrinks, so a 32 pixel postage stamp could still match full size
originals. That is a cross resolution indexing layer.
Millions of Images? Billions of Images?
Perceptual hashes (pHash) find exact resizes but break on a crop or a
tilt. FEVER matches by content, so it survives crop / rotate /
recolor. You can index at low resolution to crunch through billions of images
on a commodity hardware fleet with no GPUs in a few hours - and search at client side
native resolution for seamless retrieval. (Retrieval system patent pending)
How to get access?
Stay tuned! We are working hard to get this ready as an AWS 3rd Party
marketplace appliance. You will be able to use it commercially, to search, deduplicate,
cluster, and retrieve your photos from object storage with speed and ease - compliant with whatever your security posture could be.
Full card on Hugging Face ↗