Improve Data Loading Times

[FabianSchuetze]

Thanks for this wonderful repo. It is a pleasure to work with it.

This is not a bug or a feature request, just a simple question: Do you have any suggestions for reducing the data time during training an instance segmentation model? The masks are stored as rle and the images as jpgs.

The dataset I work with has about ~30-50masks per image. With the augmentations defined in Mask R-CNN the data time takes 0.5 (the total processing time is ~2). I set the number of data workers equal to the number of CPU cores. When working with the aggressive augmentations of RTMDet and setting poly2mask=True, the loading time is unbearable and I cannot train the network.

• Do you have any suggestions to speed up loading time?
• I still have plenty of RAM available. Could I, for example, enable some caching?
• Details about BitmapMasks & PolygonMasks #7292 seems to suggest data PolygonMasks can be more efficient. Would you advise me to change the formats?