Skip to main content
Panlabel converts between multiple annotation formats through a canonical intermediate representation (IR). All bounding boxes are represented as pixel-space XYXY (xmin, ymin, xmax, ymax) in the IR, and each format adapter handles the mapping to/from its own coordinate system. Current scope: Object detection bounding boxes only.

Format Comparison

IR JSON

Lossless canonical format

COCO

Industry standard JSON format

YOLO

Ultralytics directory format

Pascal VOC

XML-based format with Annotations/

Label Studio

Task export JSON format

CVAT

XML export for images

TFOD CSV

TensorFlow CSV format

Format Matrix

FormatPath TypeReadWriteLossiness vs IR
ir-jsonFile (.json)Lossless
cocoFile (.json)Conditional
yoloDirectory (images/ + labels/)Lossy
vocDirectory (Annotations/ + JPEGImages/)Lossy
label-studioFile (.json)Lossy
cvatFile (.xml) or directory (annotations.xml)Lossy
tfodFile (.csv)Lossy

Lossiness Matrix

What gets preserved when converting from each format to IR:
FeatureIR JSONCOCOYOLOVOCLabel StudioCVATTFOD
Dataset info
Licenses
Image metadata
Bounding boxes
Category names
Supercategory
Confidence/score
Annotations attributesPartialPartial
Images without annotations
RotationFlattened

Coordinate Systems

Each format uses different coordinate conventions:
  • IR JSON: Pixel-space XYXY [xmin, ymin, xmax, ymax]
  • COCO: Pixel-space XYWH [x, y, width, height] (top-left corner + dimensions)
  • YOLO: Normalized center-width-height [cx, cy, w, h] (0..1 range)
  • Pascal VOC: Pixel-space XYXY [xmin, ymin, xmax, ymax]
  • Label Studio: Percentage-based [x, y, width, height] (0..100 range)
  • CVAT: Pixel-space top-left/bottom-right [xtl, ytl, xbr, ybr]
  • TFOD CSV: Normalized XYXY [xmin, ymin, xmax, ymax] (0..1 range)
Panlabel handles all coordinate conversions automatically.

Next Steps

IR JSON Format

Learn about the canonical lossless format

COCO Format

Industry standard for object detection

YOLO Format

Popular training format

Pascal VOC

Classic XML-based format

Build docs developers (and LLMs) love