Lobster Detector
Sign in

Changelog

The model on duty right now is great-leap-forward: efficientnet_b0.

v0.2.0 The Great Leap Forward

Stage 2 · Fall 2026

  • A trained model replaces the random one: an ImageNet-pretrained EfficientNet-B0, fine-tuned on 5,088 photos (4,800 live animals from iNaturalist via GBIF and 288 cooked ones from Wikimedia Commons), with a second head that judges cooked versus raw.
  • Calibrated confidence (temperature scaling), and Grad-CAM heatmaps behind “Show what the model looked at”.
  • Accounts: username and password (Argon2id with a pepper), optional GitHub sign-in, lockout after repeated failures, and CSRF protection. Signing in is now required.
  • A species gallery with photos from GBIF, and community sightings.
  • Lobster Tinder.
  • “This would be great in…”: dish ideas and recipes from a local language model, with stored recipes as a fallback.
  • Rate limits, a health check, and Docker images for CPU and CUDA.

v0.1.0 Stage 1 demo

Fall 2026

  • Upload page, API, and container, with a placeholder classifier that returned random (but repeatable) results.