File size: 2,123 Bytes
9192e41 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | ---
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.
|