On-device wildlife finder and research project

Find the animal. Inspect the evidence.

Upload an enclosure photo to run the detector in your browser. The research section shows how crop choice changes recognition accuracy and cost.

Try the finder See the evidence

Photos stay on this device

Interactive finder

Search an enclosure photo

Choose a photo or use the camera. The browser downloads the detector once, keeps the image local and turns the result into a short spotting game.

spotthecreature.com · local inference ON DEVICE

Measured results

Tighter crops did not make this recognizer cheaper

Paired tests cover 466 frozen iNaturalist images across 16 species. The encoder still processed 196 patches after tight crops cut source pixels. Recognition accuracy fell.

Full image77.7%Top-1

Box crop72.5%Top-1

Mask crop67.0%Top-1

Every view196visual patches

**Recognition performance and cost**Top-1, Top-3, latency, source pixels and fixed 196-patch input.

**Paired crop effects**Bootstrap intervals and exact McNemar tests over all 466 images.

**Risk and coverage**Selective error across evidence views and the frozen linear probe.

**COD object-size sensitivity**Small camouflaged objects remain the hardest localization cases.

**Two disjoint evaluation lanes**Recognition and COD localization use different images and labels.

**Frozen supervision curve**Small amounts of labeled supervision outperform crop changes.

Frozen research profile

Validated data with a strict license boundary

Researchers verified 3,122 iNaturalist images and 6,392 COD image-mask-edge groups. Licenses restrict this collection to research, so the public detector never loads a checkpoint trained on it.

iNaturalist3,122

unique images · 16 species
2,218 / 438 / 466 split

COD / SINet6,392

image · mask · edge groups
CAMO · CHAMELEON · COD10K

ValidationPASS

all files decodable
research-only license gate

What this site publishes

The site publishes quantitative figures and four CC BY or CC0 iNaturalist examples with attribution. It omits the COD qualitative gallery because the frozen manifest lacks per-image authorship.

Validation report Figure rights ledger

iNaturalist · CC BY / CC0

Cases you can inspect

These four paired examples show context loss, mask damage, one useful crop and a failure shared by all views. The source links and licenses remain attached below the figure.

Four iNaturalist examples comparing full image, box crop, mask crop and crop with context — kclarksdnhmorg · CC BY 4.0 Brian Gratwicke · CC BY 4.0 iNaturalist photo 184618452 · CC0 keesgroenendijk · CC BY 4.0

Preprint in preparation

When Cropping Hurts

The draft presents a controlled boundary study. It measures where evidence cropping harms fixed-resolution fine-grained wildlife recognition.

01 · paired full / box / mask / context intervention02 · source-pixel vs encoded-patch accounting03 · stage-wise COD size diagnosis