lanlanlan23/XS-VID-v2

XS-VID v2

XS-VID v2 is a benchmark for extremely small object detection and tracking in videos. This release uses one canonical COCO-VID-style annotation contract across the Detection, MOT, and SOT tracks.

The annotations/, archives/, sot/, tools/, and manifests/ directories form this v2 release. All 222,924 benchmark images are packaged in 16 archives.

Download and Reproduce

Project: https://gjhhust.github.io/XS-VID/
Code and task commands: https://github.com/gjhhust/YOLOFT
Dataset mirrors: Hugging Face, ModelScope.
Checkpoints: https://huggingface.co/lanlanlan23/YOLOFT-XSVID-v2

From the YOLOFT repository, either command downloads all dataset files, checks the SHA-256 of every media archive, extracts the images and prepares portable labels/splits. ModelScope is the domestic mirror; neither command needs a token once the repositories are public. Allow about 34 GB for download plus extracted images and labels.

git clone https://github.com/gjhhust/YOLOFT.git
cd YOLOFT
python tools/xsvid/download_release.py --hub ms --destination ./downloads
# Alternative dataset mirror:
python tools/xsvid/download_release.py --hub hf --destination ./downloads
# Task checkpoints, portable TransT initialization and OSNet:
python tools/xsvid/download_release.py --hub hf --component model --destination ./downloads

See the code README for full VID/MOT/SOT inference commands and the bounded training-startup command. The corrected paper SOT checkpoint has SHA-256 c59639f868a5fdfa5474f072d657bafb072430c7386b12901e3bc1ac7a42188a. Do not substitute the early 57.09-AUC SOT checkpoint.

Layout

XS-VID-v2/
  annotations/
    train.json
    test.json
    train_dev.json
    val_dev.json
    protocols/
      paper_detection_test.json
  images/
  sot/
  tools/
  manifests/

The downloadable v2 media are divided into 16 video-aligned tar archives. Extract every file in archives/ into the same dataset root:

for archive in archives/XSVID_v2_images_*_of_16.tar; do
  tar -xf "$archive"
done

This creates the images/<video>/ tree expected by the released loaders. Verify every archive against archives/manifest.json before use.

train.json and test.json are the corrected canonical benchmark partitions. train_dev.json and val_dev.json are a development-only partition of the training data. Final benchmark models train on the full train.json and are evaluated on test.json.

annotations/protocols/paper_detection_test.json is a compact detection-only compatibility view of the evaluation annotations used for the Detection results reported in the paper. It is provided to reproduce the published table values exactly; new benchmark results should use the corrected canonical annotations/test.json.

Annotation contract

The files follow COCO-VID conventions (videos, images, annotations, and categories) and retain XS-VID fields needed by the three tracks.

• track_id is the corrected dataset-global trajectory identity used by MOT and TAO-compatible evaluation; the accompanying tracks array provides its video and category metadata.
• In test.json, paper_track_id is present only when the trajectory identity used for the submitted paper differs from track_id. The paper-compatible evaluator falls back to track_id for all other annotations.
• segment is retained when available for rich per-instance annotations. The loader falls back to segmentation if segment is absent.
• frame_index is retained for temporal ordering. The loader falls back to frame_id if frame_index is absent.
• ignore is preserved both as category ID 4 and as an annotation flag.

The sparse compatibility field avoids duplicating identity metadata on unaffected annotations while keeping canonical and paper-result evaluation in the same annotation view.

Native video loading

The legacy Detection loader requires the portable labels and split lists prepared by the download command. MOT uses canonical JSON-derived video indices; SOT uses the sequence view of the same image tree. For training, the code release provides tools/xsvid/prepare_unified_train_data.py and a bounded startup smoke command. Canonical JSON remains the source of truth; derived labels do not replace it.

YOLO-compatible labels

The canonical JSON is the source of truth. Generate a portable YOLO layout after downloading:

python tools/prepare_yolo_labels.py --root /path/to/XS-VID-v2

The command creates labels/, splits/, and xsvid.yaml. It has no symbolic-link or machine-path dependency. The default export matches the original YOLOFT recipe and retains the seven categories, including ignore. Use --exclude-ignore only for the explicit six-class no-ignore ablation. The canonical JSON remains unchanged in either case.

Use --splits-only to regenerate only the small split-list files after relocating an existing release.

Track protocols

• Detection: use annotations/test.json and the detection evaluator in the code release. To reproduce the Detection values reported in the paper exactly, evaluate the released paper predictions against annotations/protocols/paper_detection_test.json.
• MOT: use canonical track_id in annotations/test.json for new results. Use sparse paper_track_id with fallback to track_id only to reproduce the submitted paper result.
• SOT: sot/test.txt lists the official sequences, sot/image_links.json maps their frames to the shared image tree, and sot/sequences/<name>/groundtruth.txt contains the targets. sot/sequences.json records their canonical trajectory mapping. Run python tools/materialize_sot_images.py --sot-root sot --images images only when a conventional per-sequence image layout is needed.

The manifests/ directory contains SHA-256 checksums and record counts for the canonical annotations. Validate the benchmark annotations and media after download:

python tools/validate_release_annotations.py annotations/train.json --output /tmp/xsvid_train_validation.json
python tools/validate_release_annotations.py annotations/test.json --output /tmp/xsvid_test_validation.json
python tools/validate_release_media.py --root /path/to/XS-VID-v2 --output /tmp/xsvid_media_validation.json

Citation

@article{guo2026xsvid,
  title={XS-VID: A Large-Scale Benchmark for Small Object Detection and Tracking in Videos},
  author={Guo, Jiahao and Xu, Ziyang and Wu, Lianjun and Gao, Fei and Liu, Wenyu and Wang, Xinggang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026},
  doi={10.1109/TPAMI.2026.3741044}
}

License

XS-VID v2 is released under CC BY 4.0.