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 using collection filters, 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
Classification CSV	Image and whole-video classification annotations and confidence scores	CSV
COCO Object Detection	Bounding box annotations	JSON
YOLO Object Detection	Bounding box annotations	TXT labels + `data.yaml` (zipped)
COCO Segmentation Mask	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 Segmentation Mask	Video segmentation mask 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 Classifications (CSV), Video Classifications (CSV), Image Object Detections (COCO), Image Object Detections (YOLO), Image Segmentation Mask (COCO), Image Segmentation Mask (PASCAL VOC), Image Captions, YouTube-VIS Video Segmentation Masks), 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¶

The Image Filenames option exports the filenames of all samples that match the active collection filters. Apply filters in the collection view before opening the export dialog to limit the export to a specific subset.

Export sample filenames

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_csv_classifications("classifications.csv")
dataset.export().to_coco_object_detections("detections.json")
dataset.export().to_yolo_object_detections("yolo_output_dir/")
dataset.export().to_coco_segmentation_masks("segmentations.json")
dataset.export().to_coco_captions("captions.json")
dataset.export().to_pascalvoc_segmentation_mask("pascalvoc_output_dir/")

# Video dataset export
dataset = ls.VideoDataset.load()
dataset.export().to_csv_classifications("classifications.csv")
dataset.export().to_youtube_vis_segmentation_mask("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")