VPD-100K formal detection snapshot
Frozen object-detection splits for FIRE-RL. Images retain their original bytes and use numeric names with format-correct extensions.
| Split | Images | Boxes |
|---|---|---|
| Train | 82,133 | 159,587 |
| Validation | 10,271 | 15,612 |
| Test | 10,255 | 14,697 |
Annotations and validation
YOLO labels use the 33 zero-based classes in classes.txt. Train and validation labels derive from the pinned FIRE-RL training snapshot, and test labels derive from the official evaluation source. Validation and test COCO files match the exported YOLO annotations and use one-based category IDs.
A detector screened all 102,659 images for missed regions and produced 204,339 box proposals. Contextual model evidence covers all images. It matches the full current label set in 102,659 images. Of the current rows, 189,896 have evidence from an exact label match, 0 have evidence matched to the same row from an earlier label set, and 0 lack contextual model evidence for the current row.
In the final visual review round, the detector and contextual candidate totals are 1,792 and 0, respectively. The counts may overlap. Direct review of flagged regions supported corrections across 1,594 images. The correction counts are 708 additions, 931 box adjustments, and 782 removals. The count of distinct images with retained ambiguous annotations is 131 in this final review round. Direct visual review covered flagged candidates identified during screening.
Using the snapshot
From the download directory, run sha256sum -c SHA256SUMS, or shasum -a 256 -c SHA256SUMS on macOS. Extract the 27 TAR files in shards/ into the directory containing data.yaml. The shards contain matching images// and labels// paths and retain empty label files. Pass the absolute path to data.yaml to YOLO. COCO image paths are relative to the extraction directory and work with the regular FIRE-RL evaluation entry point.
Keep the original image pixels within the applicable data permissions and access scope. Project paper
- Downloads last month
- 4