TFDSv4 CARLA MOT

[davidslater]

In moving to a CARLA MOT dataset in v4, we would like to have a standard format for the labels. In particular, we are looking at an extended COCO format that looks like this:

{
"image_id": int,
"category_id": int,
"bbox": [x, y, width, height],  # not sure if it should be normalized for standard coco
"score": float,  # only used for detections, not ground truth

# The following are extended fields: 
"object_id": int,
"visibility": float, # 0 to 1
"video_id": int,
}

This is shown in more detail here (without the video_id field): https://github.com/twosixlabs/armory/blob/develop/armory/data/adversarial_datasets.py#L1013-L1024

The objects field in the builder should be a Sequence of these dicts.

This requires a number of changes:

• [ ] dataset builder: update code so that it outputs in this extended coco_format, not in the 9-dim arrays
• [ ] remove coco_format from carla-mot scenario configs (shows up 3 times)
• [ ] carla_mot baseline model: have it always output in coco format
• [ ] HOTA_metrics in metrics: default to coco format True
• [ ] carla_mot scenario: remove references to coco_format and code surrounding it
• [ ] adversarial_datasets: migrate the mot_coco_format_to_array code to HOTA_metrics as a staticmethod (since it will only be used there).
• [ ] update tests to match