Export Annotations
LabelU supports exporting annotation data in various formats, making it convenient for users to use annotation data in different scenarios.
INFO
Starting from version v1.1.0, the filenames of samples in the exported annotation files will not include prefixes. The original intention was to avoid filename duplication; now, filename conflicts will cause the upload to fail.
JSON
Supports images, videos, and audio
JSON format is the default export format of LabelU, supporting the export of annotation data for images, videos, and audio. For detailed format, refer to Annotation Format - JSON Pre-annotation.
XML
Supports images, videos, and audio
Example:
task-name.xml
<?xml version='1.0' encoding='utf-8'?>
<root>
<sample>
<id>773</id>
<folder>/Users/youname/Library/Application Support/labelu/media</folder>
<path>upload/277/57c82ae6-2018.jpg</path>
<fileName>57c82ae6-2018.jpg</fileName>
<result>
<width>321</width>
<height>481</height>
<rotate>90</rotate>
<annotations>
<object>
<toolName>rectTool</toolName>
<result>
<x>61.82131661442006</x>
<y>53.946185997910135</y>
<width>210.0919540229885</width>
<height>205.56844305120165</height>
<label>label-1</label>
<order>1</order>
<attributes />
</result>
</object>
</annotations>
</result>
</sample>
</root>
CSV
Only supports image annotations with a single annotation tool
Example:
task-name.csv
tool_name,label,label_text,x,y,width,height,attributes,order
rectTool,label-1,label-1,59.37397310283619,49.79279958727368,212.84475544654453,103.16854120486352,"0, 0, I am description-1",1
rectTool,label-1,label-1,59.56744730525572,163.03303996537298,204.22043257564826,97.59722021834268,"1, 1, I am description-2",2
COCO
Images (object detection, instance segmentation, keypoint detection)
Example:
task-name.json
{
"images": [
{
"id": 773,
"fileName": "57c82ae6-2018.jpg",
"width": 321,
"height": 481,
"valid": true,
"rotate": 90
}
],
"annotations": [
{
"image_id": 773,
"id": 0,
"iscrowd": 0,
"segmentation": [],
"area": 21958.88292252593,
"bbox": [
59.37397310283619, 49.79279958727368, 212.84475544654453,
103.16854120486352
],
"category_id": 0,
"order": 1
},
{
"image_id": 773,
"id": 1,
"iscrowd": 0,
"segmentation": [],
"area": 19931.346531170748,
"bbox": [
59.56744730525572, 163.03303996537298, 204.22043257564826,
97.59722021834268
],
"category_id": 0,
"order": 2
}
],
"categories": [
{ "id": 0, "name": "label-1", "supercategory": "" },
{ "id": 1, "name": "text-label-1", "supercategory": "" }
]
}
Mask
Images (instance segmentation)
Mask export generates a zip file that includes:
colors.json, containing color information for each category{sample-name}-segmentation.png, multi-channel mask image{sample-name}-trainids.png, single-channel mask image
Example:
colors.json
[
{
"color": "rgb(128,0,0)",
"colorList": [128, 0, 0, 255],
"trainIds": 1,
"attribute": "camera"
},
{
"color": "rgb(0,128,0)",
"colorList": [0, 128, 0, 255],
"trainIds": 2,
"attribute": "phone"
},
{
"color": "rgb(128,128,0)",
"colorList": [128, 128, 0, 255],
"trainIds": 3,
"attribute": "map"
}
]
Multi-channel mask image:
Single-channel mask image:
TIP
Grayscale image, starting from 0, each category's grayscale value + 1, supports up to 255 categories. It is difficult to distinguish with the naked eye when there are few categories.
YOLO
Images (object detection)
YOLO format export generates a zip file containing a classes.txt file and one or more txt files named after the sample names.
Example:
57c82ae6-2018.txt
1# rotate: 90
20 0.3446909580584375 0.31581641803646554 0.4425046890780551 0.32139732462574305
30 0.3361281987382117 0.6599116824752159 0.42457470389947666 0.30404118448081835
INFO
If the sample has a rotation angle during annotation, the txt file exported in YOLO format will include a comment like # rotate: 90, indicating the rotation angle of the sample.
TFRecord
Images (object detection, instance segmentation, keypoint detection)
A dataset format for the TensorFlow framework, containing images and annotation information, can be directly used for training models.
INFO
In LabelU, rotated images will also be saved as rotated in the TFRecord format export.
Pascal VOC
Images (object detection)
The exported zip file contains multiple xml files, each corresponding to the annotation information of a sample.
Example:
57c82ae6-2018.xml
<?xml version='1.0' encoding='utf-8'?>
<annotation>
<folder>/Users/gary/Library/Application Support/labelu/media</folder>
<filename>57c82ae6-2018.jpg</filename>
<path>upload/277/57c82ae6-2018.jpg</path>
<size>
<width>321</width>
<height>481</height>
<depth>3</depth>
<rotate>90</rotate>
</size>
<object>
<name>label-1</name>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>59.37397310283619</xmin>
<ymin>49.79279958727368</ymin>
<xmax>272.2187285493807</xmax>
<ymax>152.9613407921372</ymax>
</bndbox>
</object>
<object>
<name>label-1</name>
<truncated>1</truncated>
<difficult>1</difficult>
<bndbox>
<xmin>59.56744730525572</xmin>
<ymin>163.03303996537298</xmin>
<xmax>263.787879880904</xmax>
<ymax>260.6302601837157</ymax>
</bndbox>
</object>
</annotation>
TIP
The truncated and difficult fields are unique to the Pascal VOC annotation format; in LabelU annotation configuration, you need to configure the corresponding label attributes in the label configuration to export these two fields.
Configuration example:
Annotation effect:
LabelMe
Compatible with LabelMe annotation format.
NOTE
The annotation tools in LabelU are not fully compatible with the LabelMe annotation format, only supporting the export of annotation data for bounding boxes, polygons, and key points.
