---
license: agpl-3.0
pretty_name: Cubed camera tracker v1
tags:
- onnx
- computer-vision
- rubiks-cube
- keypoint-detection
- image-classification
---

# Cubed camera tracker v1

Two setup-specific ONNX models provide camera evidence for
[Cubed Core](https://github.com/KingBobJoeIV/cubed-core):

- `artifacts/alignment-classifier.onnx` classifies whether the cube is aligned
- `artifacts/face-pose.onnx` detects a face and four ordered keypoints

Cubed Core uses their outputs for alignment, pose, motion, and visible-face
color reads. The models do not emit moves, establish cube state, or complete a
decode.

## Files

| File | Purpose |
| --- | --- |
| `artifacts/alignment-classifier.onnx` | Batch-one alignment classifier |
| `artifacts/face-pose.onnx` | Batch-one face-pose model |
| `manifest.json` | Runtime roles, paths, sizes, licenses, and SHA-256 values |
| `audit.json` | Mechanical package and ONNX interface checks |
| `warm-start/camera-tracker-v1-warm-start.tar.gz` | Optional PyTorch training checkpoints |

`SHA256SUMS` covers the published payload. Version:
`camera-tracker-v1-2026-07-24`.

For the supported setup, follow Cubed Core's
[Decode guide](https://github.com/KingBobJoeIV/cubed-core/blob/main/docs/tutorials/DECODE.md).
To package or train compatible models, see
[Train tracker](https://github.com/KingBobJoeIV/cubed-core/blob/main/docs/TRAIN_TRACKER.md).
Pin an exact repository revision for reproducible use.

## Limits

These models are tied to the maintainer's cube, camera geometry, hands,
lighting, and backgrounds. They are unmeasured on unrelated setups, and no
tracker-quality, move-accuracy, or generalization claim is made. The historic
training snapshots are not included, so these exact artifacts cannot be
rebuilt from the public materials.

The warm-start archive contains PyTorch checkpoints. Loading a checkpoint uses
Python pickle and can execute code; verify its checksum and load it only in an
environment you trust. Inference users do not need it.

The published files are `AGPL-3.0-only`; see `LICENSE`. The license does not
cover unavailable historic training media.