Acquisition, pairing and image analysis of ovine carcasses, with verifiable image–animal–grade association.
Apex is a cross-platform desktop application for building a carcass image dataset and deriving objective measurements from it. Its objective is that the association between each image, the animal it came from, and the grade assigned to it is recorded at collection time and can be verified, rather than reconstructed from filenames afterwards.
A previous data-collection effort did not fail because of its machine-learning methods. It failed because the image–animal–grade association was not recorded during collection. It was reconstructed afterwards, by file order, with duplicated and illegible physical tags and a spreadsheet that had no image-name column. In Apex, a carcass is registered with a mandatory physical tag before capture, and the image is stored with a reference to it in the database.
- Collection — register a carcass (mandatory physical tag), then capture from an external camera; the image is born paired, never inferred from a filename.
- Import — bring in external photos; each becomes a carcass (or is reconciled
to one), with
sha256dedup, so nothing is orphaned. - Analysis — trained models over the batch: fat segmentation (validated, IoU ~0.92), an experimental finishing grade, and an analytical conformation convexity measure — all with automatic background removal so surface measurements are correct.
- Conformation — integral-invariant convexity of the carcass profile (leg / loin / shoulder), rendered as a blue-convex / red-concave map. Analytical, training-free, invariant to pose and lighting.
- Inter-rater grading — independent, blind grades from multiple raters, with Fleiss' κ agreement.
- Live monitor — real-time fat overlay from a camera or a video file, to help frame the carcass (background subtraction, ~real-time).
- Dashboard & export — sample progress against the target, and a dataset
export (
manifest.csv+integrity_report.txt).
| Analysis · Gallery | Analysis · Conformation |
|---|---|
 |
 |
| Project dashboard | Carcass · Images (capture + inspector) |
|---|---|
 |
 |
Wails v2 + Go (window, database, pairing) · React 19 + TypeScript + Tailwind 4 (the instrument UI, styled after NASA Open MCT "Espresso") · Python sidecars (PyTorch / OpenCV) for inference, live overlay and conformation · a single SQLite file as the source of truth.
Pre-built native binaries for macOS, Windows and Linux are attached to every release, with the model weights already bundled. Grab the archive for your platform, unpack, and run — no build tools required (the Python analysis features still need a Python with the scientific libraries below).
Prerequisites: Go 1.23+, Node 18+, the Wails CLI, and (for the
model features) a Python with torch, opencv-python, scikit-image, scipy,
rembg.
go install github.com/wailsapp/wails/v2/cmd/wails@latest
pip install torch opencv-python scikit-image scipy rembg onnxruntime
# development (hot reload)
wails dev
# if your system python3 lacks the libraries above:
export CARCASS_PYTHON=/path/to/python-with-torch
wails dev
# native build
wails buildModel weights (fat_binary.pth, direct_class.pth, eg_regression.pth,
joost_color_naming.mat) are included under model/weights/ —
see model/weights/README.md for what each one is.
A full user & reference manual (LaTeX book, with screenshots) is in
docs/, in two languages:
- English —
docs/manual.pdf(build:pdflatex manual.tex) - Português (BR) —
docs/manual-ptbr.pdf(build:pdflatex manual-ptbr.tex)
Fat segmentation and verifiable pairing are validated/guaranteed. The finishing and conformation grades are experimental estimates, not validated — the project's oracle experiment showed conformation is not reliably recoverable from a single 2D image at n=22. Apex marks these in amber throughout, and the dataset it builds is intended to support their validation. See Station 13 of the manual.
MIT — see LICENSE.