Would you be open to a PR for dataframe -> coco format?

[GeorgePearse]

I tend to use Voxel51 for most of my dataset filtering. Unfortunately the loading overhead is extreme and doesn't scale very well to large datasets.

(wherever I can replace it with pyodi I do e.g. merging annotation files)

A common workflow that I use instead is coco annotation -> dataframe -> filters -> coco annotation (normally running crude active learning to select what I should send to an annotation platform).

coco_ground_truth_to_df(ground_truth_file, max_images=200000)

I know you support the first. Would you be happy to support the latter? By this I mean can I submit a PR with some help from you to tidy it up and maybe make it more performant.