[Milestone] Refactoring data structure tracker

[shijianjian]

Bounding boxes refactoring to note about:

1. Apart from coordinates, we need to add rotated angles into our current design.
2. Similar to Keypoints not properly transformed through Augmentations #2684, there is a need to filter bounding boxes after certain transformation. This feature includes:
   • shrink the coordinates as per the image size. Then
   • remove boxes out of the image after transformations.
   • remove boxes according to a threshold (maybe a percentage of the image or the area).

Mask refactoring:

1. support list of masks as input. Support list of masks in AugmentationSequential #2421 is working on it but turns to be stale now.

Some more ideas may be suggested.

---

Note that we currently using polygon format for all the bounding boxes. Essentially, this format allows more accurate point locations after image transformations. E.g. a perspective transformation. However, it is not easy to decide how to smoothly transit from the "accurate" polygon format to the popular "XYXY(rot)" or "XYWH(rot)" formats.

[edgarriba]

• Refactor the following data structures: BoundingBox, Keypoint, Mask, Image so that follow a more type safe design.
• Keep the data structures as lights as possible with batch/non-batch support.
• Make them as much agnostic as possible to torch interface.
• Reuse internal infrastructure like Vector2, Vector3and other basic structures fromkornia.geometry`.
• Unify BoundingBox interface using a more type safe design:
  • useful pattern too: https://github.com/kornia/kornia/blob/master/kornia/contrib/object_detection.py#L23-L50
• Merge this two files into one:
  • https://github.com/kornia/kornia/blob/master/kornia/geometry/bbox.py
  • https://github.com/kornia/kornia/blob/master/kornia/geometry/boxes.py
• Keypoint can internally rely on Vector2
• Resume the Pointcloud effort Init point cloud and video augmentation module #2374

[johnnv1]

link to #2064