[mihow]

Summary

Species classifiers do better when they see the butterfly rather than the whole photo. This PR adds the research code we used to train and compare small detectors that find the butterfly in a hand-held photo, so an upload flow or a classifier can crop to it. It compares YOLOv11s, YOLO26-s and -n, RT-DETR-l, DEIMv2-S and SAHI tiling, trained on Fieldguide photos with user-drawn crop boxes and evaluated on the public Leeds Butterfly dataset. The shared pieces (data reading, splitting, evaluation metrics) go in src/localization/ next to the existing Faster R-CNN tools; the experiment scripts and write-up go in research/leps_localizer/.

The branch was rewritten into a single commit on current main. Paths and service names are now command-line arguments or environment variables, so the scripts run on any machine.

List of Changes

Change	Where
Train, evaluate, export and compare small butterfly detectors, with a methods and results write-up and a per-run results table	research/leps_localizer/
A static browser demo and a guide for running the detector inside a photo upload page	research/leps_localizer/demo/, EBUTTERFLY_INTEGRATION.md
Read the training index, repartition train/val deterministically, write YOLO datasets, and evaluate on a held-out set	src/localization/leps_data.py, repartition.py, prepare_yolo.py, eval_locked.py
Best-of-N and containment metrics for single-subject photos	src/localization/metrics.py
Unit tests for the modules above (37 pass)	tests/localization/
Optional dependency groups detection, demo and export-edge, so the base install stays light	pyproject.toml, uv.lock

Configuration comes from DATA_ROOT, RUNS_DIR, DETECTOR_DATASET_ROOT, CLASSIFIER_MODELS_DIR and, for Weights & Biases, WANDB_ENTITY.

Not included

Training images and the photo index are not part of the repository. The insect detector for light-trap photos (AnyBug) will follow in a separate PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_014Sju2W91WevdNkZhTEGyqN

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}