RaceLineCalc Track Mask
Mobile-oriented binary racetrack/asphalt segmentation model used by the automatic boundary detection feature in RaceLineCalc.
The full dataset repository is RedRatInHat/racelinecalc-track-segmentation.
Task
Input: RGB top-down / satellite-like racetrack crop
Output: single-channel logits for P(pixel is drivable asphalt track surface)
This is not a general road-scene model. It is a focused segmentation baseline for racetrack editor workflows.
Files
onnx/track_mask_lraspp_mobilenetv3_1024_v1.onnx- deployment model used by RaceLineCalc Android integration.onnx/track_mask_lraspp_mobilenetv3_1024_v1.json- model contract and benchmark metadata.pytorch/checkpoint_best.pt- best PyTorch checkpoint from the training run.training/- run config, split, metrics and test metrics.assets/racelinecalc_track_mask_predictions_3x3.png- visual example of model predictions.
Model Contract
input_name: image
input_shape: 1 x 3 x 1024 x 1024
input_dtype: float32
normalization.mean: [0.485, 0.456, 0.406]
normalization.std: [0.229, 0.224, 0.225]
output_name: track_logits
output_shape: 1 x 1 x 1024 x 1024
output_dtype: float32
output_semantics: logits for P(pixel is drivable asphalt track surface)
Apply sigmoid and a threshold outside the graph. RaceLineCalc uses the mask as input to a boundary tracing postprocess.
Training Summary
- Architecture: LR-ASPP + MobileNetV3 Large
- Dataset: RedRatInHat/racelinecalc-track-segmentation
- Samples: 120
- Image size: 1024 x 1024
- Best epoch: 5
- Validation IoU: 0.9433
- Test IoU: 0.9370
- Test Dice: 0.9675
- Test Boundary F1: 0.9395
Android Benchmark
Measured on Ulefone Armor 28 Ultra, Android 15, ONNX Runtime Android with NNAPI CPU disabled and FP16 enabled:
- Warmup runs: 5
- Benchmark runs: 30
- Median latency: 752.16 ms
- P95 latency: 761.45 ms
Related Dataset
The full dataset is published here:
RedRatInHat/racelinecalc-track-segmentation
Provenance
The model was fine-tuned for RaceLineCalc automatic track boundary detection. The architecture is based on torchvision LR-ASPP MobileNetV3 Large, then adapted to binary segmentation and exported to ONNX for mobile runtime evaluation.
License
License is marked as other. Commercial use of this model and its weights is not allowed without a separate written agreement with Red Rat In Hat.