Other annotation formats than Json for memory efficiency

[chriss2401]

Describe the feature
I have successfully trained a mask rcnn network by using mm detection on a relatively small custom dataset. I was wondering whether it is possible to use a custom dataset that is very large without having to write the annotations in a json file, since json in general isn't made for handling such large files.

Motivation
Ex1. It is inconvenient when I need to serialize a large amount of my custom annotations to json. Many times I get out of memory errors when trying to serialize these annotations to a json file (coco format).
Ex2. Json is not made for such large datasets, therefore the resulting annotations can take up a lot of space.

[ZwwWayne]

Thanks for your kind suggestion. For now, we use JSON for convenience. If you want to use other formats, you may consider re-implement the logic of the dataset. We will consider the efficiency issue in the future.