If we have a dataset of polygons that annotate our target class, but there are some gaps with no annotations present, we can still use this data to create a COCO dataset.
We can define a function to search for squares in the vector polygon dataset where a defined percent of the area is annotated. This function can then save the raster that falls beneath this square. These square tiles can then be passed through the gis2coco workflow as normal.
This would likely be easiest as a slightly modified version of the geojson2coco.py script.
If we have a dataset of polygons that annotate our target class, but there are some gaps with no annotations present, we can still use this data to create a COCO dataset.
We can define a function to search for squares in the vector polygon dataset where a defined percent of the area is annotated. This function can then save the raster that falls beneath this square. These square tiles can then be passed through the gis2coco workflow as normal.
This would likely be easiest as a slightly modified version of the geojson2coco.py script.