Read the image in local browser memory.
Live neural network · no photo uploads
Field intelligence, running on your device
Every field mark, considered.
A fine-grained vision model trained on 48,562 NABirds photographs to recognize 555 North American bird classes — and show you when it is not sure.
0bird classes
0training photographs
0top-5 test accuracyright class appears among five guesses
0calibration error
From photograph to shortlist
A serious model in a static page.
GitHub Pages serves the interface and an 18 MB ONNX model. Your browser performs every calculation locally.
Resize, center-crop, and normalize to 224 × 224.
Run EfficientNet-B0 across all 555 learned classes.
Return ranked probabilities and an uncertainty verdict.
Your photo stays yours.It is never posted to GitHub, an API, or an analytics service. Closing the tab clears the in-memory prediction session.
Read the numbers honestly
On the untouched 24,633-image test split, the exact class was the first guess 52.0% of the time, inside the top three 71.1% of the time, and inside the top five 78.0% of the time. Top-5 is not a 78% guarantee for one photo; it means the correct answer appeared somewhere in the five-name shortlist across the full test set.
Field guide
Give the model a fair look.
Fine-grained identification depends on details: bill shape, eye rings, wing bars, posture, and plumage.
- 01
Frame one bird
Use a focused photo where one bird fills a useful part of the image. Side and three-quarter views expose more field marks.
- 02
Upload privately
Drop JPEG, PNG, or WebP files on the Identify tab. Batch them if you like; each stays in your browser.
- 03
Read the badge
Green is the strongest verdict. Amber means compare alternatives. Red means unsupported species, an unusual image, or low confidence.
- 04
Use the shortlist
Look-alikes are hard. Treat the ranked alternatives as a birder’s shortlist, not five independent guesses.
Photograph do’s and don’ts
- One bird, reasonably large in frame
- Natural light with plumage in focus
- Side or three-quarter profile
- Original image with minimal overlays
- Flocks with several possible subjects
- Silhouettes or severe backlighting
- Distant birds occupying very few pixels
- UI screenshots, watermarks, or heavy filters
What the settings mean
Alternatives
Choose three, five, or ten ranked candidates. Five balances useful context and readability.
Confidence threshold
Below this probability, Fieldmark flags a result as possibly unknown. It is a caution threshold, not an open-set detector.
Technical details
Expose entropy, top-two margin, calibration temperature, runtime, and inference time for a deeper read.
Closed-set model
The network must choose among the 555 classes it learned. Geography, season, age, and sex are not separate inputs.
The live model
Identify a bird.
Preparing the private browser model…
Drop bird photographs here
or browse your files · JPEG, PNG, WebP · up to 20 MB
No sightings yet
Upload a photograph and its calibrated shortlist will appear here.