Seal or Sea Lion?
Upload a photo and a 17 MB MobileNetV3, running as ONNX in your tab, says whether it’s a true seal, an eared seal (sea lions and fur seals, who are on the same team whatever their name says), a walrus, or none of the above. Trained on iNaturalist photos; the photo never leaves your device.
- PyTorch
- timm
- ONNX
- onnxruntime-web
- iNaturalist API
- CLIP
Drop a photo here
or paste one, or pick one below. It never leaves your device: the model runs in this tab.
model not loaded yet
No photo handy? Try one of these
Photos by Brian Gratwicke, John Barkla, patrickc, Chelle Temple-King on iNaturalist, CC BY.
Verdict
Waiting on a photo.
Seal, sea lion, fur seal, walrus, or none of the above. I’ll tell you which family and how sure I am.
Why three families, not two teams
“Seal vs sea lion” is the popular question, and it’s the wrong one: fur seals are otariids, which puts them on team sea lion no matter what their name says. Pinnipeds actually split into three families, so that’s what the model predicts. The answer is always right at that level, and the question mark in the title is there so the result can correct the premise.
- True sealPhocidae
- harbor, grey, elephant, leopard and monk seals, among others. No ear flaps, short clawed front flippers, and it moves on land by wriggling on its belly.
- Eared sealOtariidae
- every sea lion and every fur seal. Visible ear flaps, long front flippers, and it can rotate its hind flippers to walk.
- WalrusOdobenidae
- just the walrus. Tusks, a moustache of whiskers, and sheer size.
How good is it, honestly
It was trained on iNaturalist photos of wild animals (CC0 and CC BY only, about 8,803 of them) and scores 88.5% on the pinniped photos in a held-out test set, rising to 92.9% on the photos it’s confident about, which is why anything under 85% gets an “unsure” instead of a verdict. Fur seals, the case the app exists for, land at 91.3%. A lone head in the water and walruses without visible tusks are the hard cases; a clear, close-up photo is the easy one. The 95% I aimed for was not reached, and the README has every number, including the ones I’d rather it didn’t.
MobileNetV3-Large fine-tuned at 288 px, exported to a 16.82 MB ONNX file and run with onnxruntime-web (WASM) in your tab. The photo is decoded, cropped and scored locally; nothing is uploaded, and there is no server on the other end to upload it to.