Omatrack telemetry model bundle
Selected local-inference models for the Omatrack 1.8.3 patch release. Release builds download an immutable revision and verify every file hash; the resulting applications include the weights for offline operation. No video is uploaded. This is a custom ONNX contract, not a drop-in Transformers pipeline.
| File | Purpose | Size |
|---|---|---|
gauge-reader.onnx |
Proven four-crop value reader | 2,213,746 bytes |
detectors/tiny-v2/gauge-detector.onnx |
Lightweight general proposal discovery | 966,419 bytes |
detectors/aim-large-v1/gauge-detector.onnx |
Higher-accuracy known-AiM proposal discovery | 112,025,875 bytes |
The reader is byte-identical to the existing tobil/omatrack-telemetry-reader v1.0.0 artifact. The original repository and revision are unchanged.
Intended routing and safeguards
Use the larger detector only after the current full source image independently matches the reviewed 1920x1080 orange AiM layout. Use the lightweight detector for other layouts and as a failure fallback. Do not infer layout from file extension or filenames. Source dimensions/display transforms must be verified before boxes are projected onto playback.
All learned detections are proposals, not calibrated probabilities or reader approval. The reader outputs gear, the displayed stint counter, brake visible-fill percent and throttle visible-fill percent only for its reviewed crops. It does not measure physical brake pressure, GPS, lap classification, arbitrary fonts or every other detected widget. Unsupported crops stay unknown. Native telemetry takes priority. User confirmation precedes extraction.
The detector tensor/preprocessing/postprocessing contract is documented in each
contract.json: RGB float32 NCHW [1,3,384,640], full-source half-pixel bilinear
resize without antialiasing, then five stride-four prediction heads. Read each
contract in full; class ordering, sigmoid/offset math, thresholds and NMS matter.
Evidence and limitations
These are known-style, session-held-out development checks, not an unseen-layout or broad no-HUD certification. Private evaluation images are not included.
- Original eight full frames: lightweight precision77.5% / recall69.4%; larger precision73.5% / recall82.6%. The larger model is NOT a global replacement.
- On the original VBOX subset, larger precision fell to58.2% from75.3%.
- Additional eight exhaustively reviewed AiM frames: larger precision97.8% / recall99.3%, versus lightweight83.3% /66.2%.
- Additional eight partial VBOX frames: larger annotated-target recall85.5%; precision is unscored. Estimated pedal-track extents were withheld from exact IoU scoring, not treated as absent.
- Larger model produced two false proposals on16 additional reviewed no-gauge crops. Such crops are not full no-HUD recordings.
- ARM-server CPU measurements: the larger model costs about388ms inference / 410ms including file decode, preprocessing and box decode with two ORT threads. One-thread native processing is around0.8s. Discovery runs only every few seconds on a worker; this is not a claim of5Hz whole-frame detection.
The backbone, annotation coverage and training recipe changed together; the comparison is not an isolated model-capacity experiment. The more heavily framing-randomized value-reader candidate failed clean/extreme-framing gates and is deliberately NOT included.
Privacy and rights
Only selected ONNX models, their companion contracts/metadata, checksums and scoped upstream notices are published. No private frames, telemetry, labels, training checkpoints, outputs or real-image fixtures are redistributed. A separate license for the task-specific weights remains unspecified; see NOTICE.md. Do not infer a blanket MIT/Apache/BSD grant for the task weights or source footage.
Application code and runtime documentation: tobi/omatrack.