Teaching a model which way is up.

The problem

A camera that has been knocked, rolled or mounted upside down still reports itself as healthy. Existing orientation models were trained mostly on tidy photographs, and camera frames are compressed, dim and noisy.

The Wild Ducks part

Wild Ducks did the model engineering with the ORTUS AI team: training data with a recorded source and licence for every image, a benchmark protocol that does not leak the answer, abstention fixed on calibration data, and runtimes for CPUs, NVIDIA GPUs, Apple silicon and web browsers.

What was built

• Six tiers, from Pico, small enough to run in a web browser, to Max for the hardest frames, with cascades in between.
• Training and testing under CCTV-style degradation, so the result holds up on dim, blurred and compressed frames.
• An explicit abstention for views with no “up”, such as open sky or bare ground, with thresholds fixed once on calibration data and left alone.
• A sealed holdout, opened once after the weights were frozen, and comparisons scored from each model’s public release.

Outcome

Released on 30 September 2026 under Apache-2.0: code, weights for all six tiers in ONNX and Core ML, benchmark scripts and a technical report. The demo above runs the smallest tier on your device.

• Release write-up and film · ORTUS AI
• Code and technical report · GitHub
• Weights for all six tiers · Hugging Face
• Python package · PyPI