PASCAL VOC format

The Pascal VOC format exports bounding boxes from image and video projects, as one XML file per image.

Only bounding boxes are exported. Polygons, semantic segmentation, points, lines, pose estimation and classification are left out. For polygons and semantic segmentation, use YOLO or COCO.

Archive structure

export.zip
├── README.kili.txt
├── images/
└── labels/
    ├── image_1.xml
    └── image_2.xml

All jobs are grouped in the same files: Pascal VOC has no per-job layout.

For video assets, each frame that holds an annotation gets its own image and XML file, and a video_meta.json file lists the frames of each video.

Label files

Each XML file describes one image: its file name and size, then one object element per bounding box, with its class name, its job, and its corners in pixels (xmin, xmax, ymin, ymax). Boxes are kept inside the image.

<annotation>
   <folder></folder>
   <filename>image_1.jpg</filename>
   <path></path>
   <source>
      <database>Kili Technology</database>
   </source>
   <size>
      <width>1920</width>
      <height>1080</height>
      <depth>3</depth>
   </size>
   <segmented>0</segmented>
   <object>
      <name>Car</name>
      <job_name>VEHICLES</job_name>
      <pose>Unspecified</pose>
      <truncated>0</truncated>
      <difficult>0</difficult>
      <occluded>0</occluded>
      <bndbox>
         <xmin>120</xmin>
         <xmax>480</xmax>
         <ymin>340</ymin>
         <ymax>610</ymax>
      </bndbox>
   </object>
</annotation>