VisDrone2019 DET — training TSV bundle

Derived from the AISKYEYE team at Tianjin University's VisDrone dataset. Train only: 6,471 images, 343,205 target boxes and 10,345 ignored rows. Validation and test are not packaged. Exhaustive validation passed with zero filename, original-byte, or decoded-RGB overlap against the 548 validation images. Two duplicate-image pairs and one zero-height car target remain unchanged. The target is in 9999985_00000_d_0000020.jpg, source row 11 (zero-based), xyxy [611, 158, 615, 158]. A training selection must exclude this invalid target; the immutable export retains it for provenance. This uses the same indexed TSV storage and standalone loader as the project's RefCOCOg training export. It is not the separate RefCOCOg Parquet/image-shard release and does not use datasets.load_dataset.

Loading

Download the repository files, install Pillow, and run from the downloaded root:

from loader import RawTSVDataset
data = RawTSVDataset('visdrone')
sample = data.decode(data[0])
image = sample['image']
boxes = sample['annotation']['boxes']

Image rows contain a byte offset and base64 of the original image bytes. Annotation rows contain an image byte offset and JSON. .lineidx files contain byte offsets. The positive index selects images with at least one target. Every original train image is retained, without re-encoding or EXIF rotation.

Targets and ignored regions

One annotation row per image. boxes contains absolute pixel xyxy boxes for source score=1 and category IDs 1–10. Class names are pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor. Pedestrian and people remain distinct. Configure a grounding reader with ori_box_format="xyxy". ignored_regions retains score=0, category 0, and category 11 entries; these are not positive targets or negative training queries. Truncation and occlusion attributes are preserved. Coordinates are not clipped, normalized, or quantized.

The schema is rex-tsv-visdrone-detection-v1. source_annotations preserves all eight original integer fields per row; source_annotation_base64 preserves the exact annotation-file bytes. Image/annotation identities and dimensions are included. Follow the official DET rules for ignored areas during evaluation. Runtime model adapters must choose their prompt, answer format, ignored-area policy, and dense-image token budget. A generic box reader does not automatically implement ignored-area evaluation.

Validation and publication status

manifest.json records source archive and payload SHA-256 checksums. validation.json records exhaustive image decoding, exact source-byte comparisons, annotation mappings, byte indexes, category counts, geometry issues, and exact train/validation image overlap. Exact overlap checks do not establish that scenes or adjacent video frames are independent. No model training was performed. ready_for_training: false means a model-specific adapter is still required.

The supplied ZIPs contain no license file, and the official dataset README does not specify redistribution terms. No third-party license claim is adopted here. Publication readiness remains blocked until applicable redistribution terms are established. Source attribution does not itself establish permission.

Please cite: Zhu et al., “Detection and tracking meet drones challenge,” IEEE TPAMI 44(11), 7380–7399, 2021, as requested by the source repository.