dictvv/voc-yolo

VOC dataset (YOLO format, ready-to-use)

Pre-processed Pascal VOC 2007 + 2012 dataset, already converted to YOLO label format. Contains both images (JPEG) and labels (YOLO txt) packaged in a single uncompressed tar (~2.9 GB).

Usage

pip install huggingface_hub
hf download dictvv/voc-yolo voc.tar --local-dir /workspace/datasets --repo-type dataset
tar -xf /workspace/datasets/voc.tar -C /workspace/datasets/
rm /workspace/datasets/voc.tar

After extraction /workspace/datasets/VOC/ is ready to use with ultralytics:

from ultralytics import YOLO
m = YOLO('yolov8n.yaml')
m.train(data='VOC.yaml', ...)

Directory structure

VOC/
├── images/{train,val,test}{2007,2012}/  (JPEG images, moved out of VOCdevkit)
│   └── VOCdevkit/                        (original extract, annotations + image sets)
└── labels/{train,val,test}{2007,2012}/  (YOLO format .txt, one per image)

Splits

Following ultralytics VOC.yaml convention:

• train (16551 images): train2012 + train2007 + val2012 + val2007
• val (4952 images): test2007
• test (4952 images): test2007

Source

Original zips from ultralytics GitHub Releases (which mirror Pascal VOC official). Label conversion uses the same XML to YOLO logic as ultralytics VOC.yaml download script.