Dataset consolidation, validation, export, training, and evaluation pipeline for underwater robotics perception (RoboSub), starting with the gate task.
Status: MVP. The full pipeline runs end to end on a tiny synthetic fixture. Real-data importers are not yet implemented.
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytestTested with Python 3.11. requirements.txt has the exact versions used.
Prepare a dataset (generate fixture, import, validate, dedupe, split, export YOLO Detect):
python -m dataset_tools.prepareTrain (1-2 epoch smoke test):
python training/train.py configs/smoke_detect.yamlOutputs go to runs/smoke/ (gitignored).
dataset_tools/make_fixture.py: tiny fake dataset with known boxes and fake scene IDsdataset_tools/importers/: converts a source into the common manifest (data/manifest.json)dataset_tools/validators/: schema values, required fields, pixel-bounds checksdataset_tools/splits/: hash dedupe, scene-aware train/val/test split with a leakage checkdataset_tools/exporters/: YOLO Detect layout anddataset.yamltraining/,configs/: config-driven YOLO training with a fixed seed
Missing values are stored as null, never guessed. Records without a usable
label are excluded from an export instead of getting invented labels.
- Importers for the real synthetic generator output and reviewed SAM labels
- YOLO Pose and Segment export
- Visual overlays for QA
- Evaluation (precision/recall, keypoint error, failure cases, run metadata)
- Frozen real-world evaluation split
- Shared-storage (OCF) artifact workflow
- CI
- Final schema