[Ai-1477]-darwin-v2-export-fix

[ChristofferEdlund]

Problem

Recent changes to darwin introduced local datasets loading of only darwin_v2 format.

When converting video datasets to images by using the split function split_video_annotations

darwin-py/darwin/dataset/remote_dataset.py

Line 143 in 79fd606

def split_video_annotations(self, release_name: str = "latest") -> None:

there are usage of a darwin export will not generate v2 annotations, instead it generates this json (assumably v1)

darwin-py/darwin/exporter/formats/darwin.py

Line 17 in 79fd606

def build_image_annotation(annotation_file: dt.AnnotationFile) -> Dict[str, Any]:

    return {
        "annotations": annotations,
        "image": {
            "filename": annotation_file.filename,
            "height": annotation_file.image_height,
            "width": annotation_file.image_width,
            "url": annotation_file.image_url,
        },
    }

Leading to our dataloaders not being able to load any data that has been converted from video now that v2 format is enforced.

Solution

Updating the darwin v2 export function build_image_annotation to generate valid v2 image annotations.
It's only used in that one place above currently so changing it should fix our problem and not introduce any unwanted issues.

Changelog

• Updated split_video_annotations convertion so it generates valid per-image darwin v2 annotations.
• Added option to select if one wants to load empty annotations in LocalDataset with keep_empty_annotations arg (default is False)

[linear]

AI-1410 darwin-v2-export-fix

[Nathanjp91]

Some stuff to look at.

[linear]

AI-1477 Export Fix for video annotations