[COCO Dataset improvement]

[cregouby]

Current Situation

coco_detection_dataset is carrying object detection and object segmentation at the same time. This comes with

• the need to carry a huge 500MB+ $annotation object hosting all annotations for both tasks along the whole dataset lifetime in memory (i.e. forever)
• a complex documentation for the dataset that is confusing to run one task or the other, knowing that the two task will never run together (as they target different model architecture)
• the downloaded archive and ann_zip are stored at root folder on the rappdirs::user_cache_dir("torch") without using a prefix = but have filename not explicit to the coco dataset, so end-user will have to deal with an hardly un-identified val2014.zip 6GB files on disk when reaching the "disk full" exception...

Suggestion

As the two tasks have no common DNN implementation, we could easily (through instantiation)
split the coco_detection_dataset() into

• coco_detection_dataset()
• coco_segmentation_dataset()

In the website (see pkgdown.yml) we should also separate classification datasets form the non-classification datasets.

That would allow much more straightforward example, and lower memory footprint of $annotation. That would also ease #170

• archive and ann_zip should be moved into or prefixed by a /coco folder for better identification in the cache folder