Export¶

Export lets you save annotations and file paths from your dataset in common formats. Exports can be triggered from the GUI or through the Python API. You can export an entire dataset or a filtered subset based on tags, or other criteria when using the Python API.

The following export formats are supported:

Format	Content	File type
Sample Filenames	Absolute file paths of selected samples	TXT
COCO Object Detection	Bounding box annotations	JSON
COCO Instance Segmentation	Segmentation masks with bounding boxes	JSON
COCO Captions	Image captions	JSON
Pascal VOC Semantic Segmentation	Per-pixel class masks	PNG masks + class map
YouTube-VIS Instance Segmentation	Video instance segmentation tracks	JSON

Export in the GUI¶

Open the export dialog from the Menu button in the top-right corner of any grid view and select Export. The dialog shows a dropdown with all available formats for the current dataset.

Export format dropdown

Exporting annotations¶

For annotation formats (Image Object Detections, Image Instance Segmentations, Image Semantic Segmentations, Image Captions, YouTube-VIS Video Instance Segmentations), select the format and click Download. A zip file is created if the export includes multiple files. The export includes all samples in the dataset.

Export object detections

Exporting sample filenames by tag¶

The Image Filenames option supports tag-based filtering. Tick Inverse selection to export samples that do not have the selected tag.

Export filenames by tag

Export in Python¶

Export a dataset¶

Call dataset.export() to get an export object, then call a format-specific method on it. The export methods depend on the dataset type, see Dataset export reference for details.

import lightly_studio as ls

# Image dataset export
dataset = ls.ImageDataset.load()
dataset.export().to_coco_object_detections("detections.json")
dataset.export().to_coco_instance_segmentations("segmentations.json")
dataset.export().to_coco_captions("captions.json")
dataset.export().to_pascalvoc_instance_segmentation("pascalvoc_output_dir/")

# Video dataset export
dataset = ls.VideoDataset.load()
dataset.export().to_youtube_vis_instance_segmentation("youtube_vis.json")

Export a filtered subset¶

Pass a DatasetQuery to export() to export only matching samples. Queries support filtering, ordering, and slicing — see Search and Filter for the full query API.

import lightly_studio as ls
from lightly_studio.core.dataset_query import ImageSampleField

dataset = ls.ImageDataset.load()

# Export only samples taller than 500 pixels
query = dataset.query().match(ImageSampleField.height > 500)
dataset.export(query).to_coco_object_detections("tall_images.json")