ghmhhjvb/ardra-drone-detection-v2

Ardra drone detection v2

Browse this dataset by object class first, then by train, val, or test. The five image folders at the repository root are:

• marker_flag/
• instrument_box/
• hammer/
• short_tripod/
• hard_negatives/ — verified target-free images; there is no fifth COCO object category.

Within each class, use train/canonical/images/ for clean training images and train/augmented/images/ for generated training variants. Clean validation and final test images are in val/canonical/images/ and test/canonical/images/. Drone variants are in val/augmented/drone/images/ and test/augmented/drone/images/; additional test conditions are under test/augmented/stress/<category>/images/.

Every image file exists once in this repository. There are 24,245 distinct images: 3,360 canonical and 20,885 augmented. The paths changed from the GitHub v2 release, but the image bytes and SHA-256 values did not. metadata/path_migration.csv maps every old path to its new path; metadata/image_index.csv lists the current paths and hashes.

Bounding boxes and training

COCO boxes remain in annotations/. All file_name values are relative to the repository root, so set the COCO image root to the downloaded dataset directory. The most convenient annotation files are:

Purpose	COCO file	Images
Train, clean	annotations/train/canonical/instances.json	2,341
Train, augmented	annotations/train/augmented/instances.json	14,046
Validate, clean independent scenes	annotations/val/instances.json	534
Final test, clean independent scenes	annotations/test/instances.json	485

The 18 separate source-split COCO files remain under annotations/{train,val,test}/... for clean, drone, and stress-specific evaluation. If your trainer requires a single train COCO file, run python scripts/merge_train_coco.py . train_instances.json after downloading; it combines the two train files with fresh unique IDs. COCO category IDs are 0 marker_flag, 1 instrument_box, 2 hammer, and 3 short_tripod. Hard negatives have no boxes. See TRAINING_GUIDE.md.

Source-scene groups and all their variants stay within a single train/val/test partition. Augmented val/test images are robustness checks of existing scenes, not additional independent evaluation scenes. Training cannot be guaranteed free of overfitting or underfitting; assess per-class results and learning curves.

The frozen QA review recorded 3,210 fully verified source annotations, 13,789 still needing full adjudication, 39 confirmed corrections, one pending fix, and 73 missing-instance exceptions. Existing boxes on unadjudicated images are retained. Confirmed missing objects without defensible exact boundaries are documented but remain unboxed. This is a usable dataset, not a claim that every label is perfect.