TFDS object detection to coco format

[davidslater]

The object detection datasets are not in standard COCO format. We should move them to output in standard COCO format.
This file is a good model:
https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/object_detection/coco.py

See here for the spec (under 1. Object Detection):
https://cocodataset.org/#format-data

We will want a format similar to:

'objects':
              tfds.features.Sequence({
                  'id': np.int64,  # coco has unique ID for each detection. We may omit this field, I think.
                  'image_id': np.int64,
                  'area': np.int64,
                  'bbox': tfds.features.BBoxFeature(),
                  'label': tfds.features.ClassLabel(num_classes=80),  # this maps down to the set of categories actually used
                  'iscrowd': np.bool_,
              }),

This requires ensuring that each dict only is a single detection.

One challenge is that the internal format for tfds does not necessarily correspond to coco format. For instance the BBoxFeature uses normalized floats [xmin, ymin, xmax, ymax] instead of coco's unnormalized floats [x, y, width, height].

Currently, some datasets have different names (e.g., xview uses boxes instead of bbox, carla datasets also, I think).
Also, some datasets (xview again) have a dict of arrays instead of a sequence of dicts.

We should be very explicit about what we expect / allow for use in Armory, and be consistent.

[lcadalzo]

From discussion with @davidslater: the intention is to continue having raw datasets build in the tf format of normalized [y1, x1, y2, x2], but in preprocessing instead of converting to pytorch format of unnormalized [x1, y1, x2, y2] (which is what we do currently), we'll convert to the coco format of unnormalized [x1, y1, width, height]. This should enable use of pycocotools