Detector-domain baseline package

This package contains exactly six locked-protocol baseline jobs: RT-DETR, Faster R-CNN R50-FPN, and SSDLite320-MobileNetV3-Large, each trained once on bright (D_B) and raw dark (D_L) LOD images.

The six notebooks are in notebooks/. Attach this directory as a Kaggle dataset, attach the LOD image dataset, edit only the first notebook cell, and choose Run All. Each run verifies the paired 1647/183/400 split, smoke-tests model/loss/backward/checkpoint reload, trains for 50 epochs, selects solely on the matching-domain validation mAP50-95, evaluates the required locked test sets, and emits detector_results.zip.

All jobs use zero-based source labels for the eight frozen classes: car, motorbike, bicycle, chair, diningtable, bottle, tvmonitor, bus. Reported AP uses the common pycocotools COCO bbox evaluator (maxDets=100), rather than framework-specific metrics.

Outputs contain config.yaml, resolved_config.yaml, best_checkpoint.pt, last_checkpoint.pt, train_log.csv, val_metrics.json, test_metrics.json, and run_metadata.json. D_B evaluates both bright and raw-dark test; D_L evaluates raw-dark test only.

Local command

python scripts/verify_dataset.py --dataset-root /path/to/lod --labels-root labels --package-root .
python scripts/train_detector.py --config configs/rtdetr_db.yaml --dataset-root /path/to/lod --labels-root labels --output-dir outputs/rtdetr/D_B --smoke
python scripts/train_detector.py --config configs/rtdetr_db.yaml --dataset-root /path/to/lod --labels-root labels --output-dir outputs/rtdetr/D_B

Pretrained initialization is recorded in each YAML. Within each detector architecture, D_B and D_L differ only in image domain/manifests.