feat(deepen_to_t4_segmentation): add deepen_to_t4(object_ann.json) converter for 2d segmentation

[Shin-kyoto]

Description

• Add deepen_to_t4 converter for 2d segmentation.
• TODO: add tests

How to review

• Please check Notes for reviewer, and give me your advice.

How to test

test data

• To be filled

test command

Reference

Notes for reviewer

The following are the challenging parts of the implementation:

• Since ObjectAnn always requires a 2D bounding box, a dummy value is being inserted for the 2D bbox.

  • For 2D detection, a dummy mask encompassing the entire image is also being used. This results in an unnatural aspect of the overall implementation.

• The mask argument of _convert_to_t4_format is of type List[Dict[int, str]], which is insufficient for storing masks per frame and instance.

  • It is currently used for 2D detection as mask[sensor_id][frame_index].
  • As a result, the interface is not expandable for 2D segmentation.

• In this implementation, as a last resort, I replaced the process where anno_two_d_box is created with generating RLE instead.

  • However, creating RLE requires height and width, which forced me to choose between creating an instance of the nuImages class or using the size from the mask. I opted for the latter, but this approach involves generating a mask that encompasses the entire image, solely to obtain the size, which is not necessary for 2D segmentation and is an inelegant solution.

• This code works well if I use polygons[0] in line 306 of annotation_files_generator.py. But I do not know the reason 🤔