XS-VID v2 Download and Evaluate

Download the dataset

The YOLOFT downloader uses the Python standard library; no hub SDK installation is needed. It verifies archive SHA-256 checksums, extracts images and prepares labels, split lists and the Detection YAML bridge.

Allow about 34 GB for downloads, plus space for extracted images and labels. Replace ./downloads with your destination directory.

git clone https://github.com/gjhhust/YOLOFT.git
cd YOLOFT
python tools/xsvid/download_release.py --hub ms --destination ./downloads

Alternatively, use Hugging Face:

python tools/xsvid/download_release.py --hub hf --destination ./downloads

Dataset mirrors: ModelScope / Hugging Face.

Model checkpoints

YOLOFT / Unified YOLOFT weights: vid.pt, unified_mot.pt, unified_sot.pt.

python tools/xsvid/download_release.py --hub hf --component model --model-repo lanlanlan23/YOLOFT-XSVID-v2-weights --destination ./downloads

Prepared dataset layout

downloads/XS-VID-v2/
  annotations/
  archives/
  images/
  labels/
  splits/
  sot/
  tools/
  manifests/
  legacy_yoloft.yaml

The dataset contains 222,924 images in 16 media archives. Image extraction and label preparation are automatic with the dataset command. Background frames intentionally have no label file.

Evaluation protocols

See the Results for track comparisons.

Detection: --detection-protocol paper selects annotations/protocols/paper_detection_test.json; canonical selects annotations/test.json.
MOT: paper protocol uses paper_track_id (fallback: track_id); canonical uses track_id.
SOT: official 272-sequence test list and target trajectories.

Task-specific inference, evaluation and training commands are in the YOLOFT documentation.

Versions and existing downloads

Record the code commit, model release version and dataset revision. HF defaults to main and ModelScope to master; pass --revision YOUR_RECORDED_REVISION to pin the dataset revision.

Rerunning verifies and reuses downloaded files. Failed transfers are retried. Use --offline to verify and prepare an already complete download.

SDK alternatives and manual validation

These optional SDK commands download repository files only. Extraction and label preparation must be run separately.

python -m pip install -U huggingface_hub
python -c 'from huggingface_hub import snapshot_download; snapshot_download(repo_id="lanlanlan23/XS-VID-v2", repo_type="dataset", local_dir="./downloads/XS-VID-v2")'

python -m pip install -U modelscope
python -c 'from modelscope.hub.snapshot_download import dataset_snapshot_download; dataset_snapshot_download(dataset_id="lanlanlanrr/XS-VID-v2", local_dir="./downloads/XS-VID-v2")'

Verify archive hashes against archives/manifest.json before extraction. From the YOLOFT code root:

for archive in ./downloads/XS-VID-v2/archives/XSVID_v2_images_*_of_16.tar; do
  tar -xf "$archive" -C ./downloads/XS-VID-v2 || exit 1
done
python tools/xsvid/prepare_legacy_layout.py --data-root ./downloads/XS-VID-v2 --detection-protocol paper --regenerate-labels

Optional validation runs from the downloaded dataset root. Keep the complete tools/ directory and its dependencies:

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

These validators check annotations and referenced file presence; the media validator does not hash image pixels.

Licenses

Consult the dataset license, YOLOFT code notices and applicable pretrained-weight licenses.