1114 rle support for coco

[emSko]

Description

PR adds support for the RLE format in the COCO dataset. Based on #1114 .

It required implementing RLE encode/decode functions and extending the coco_annotations_to_detections/detections_to_coco_annotations functions.

Type of change

Please delete options that are not relevant.

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

How has this change been tested, please provide a testcase or example of how you tested the change?

All functionalities are covered with unit tests + notebook checks the functionalities on larger images.

Any specific deployment considerations

None

Docs

• Docs updated? What were the changes: docs/datasets/utils.md

[CLAassistant]

All committers have signed the CLA.

[SkalskiP]

Hi @emSko 👋🏻 Could you accept CLA? Without this, we won't be able to merge your PR.

[SkalskiP]

Let's split the work on this PR into three parts:

• Implementation of rle_to_mask and mask_to_rle. Please write valid docstrings in Google format for both methods. For example, take a look at other documented functions in our repository here. The documentation should describe the input and output arguments, as well as provide a simple usage example. Regarding the tests, create several cases: when the image is entirely zeros, when the image is entirely ones, when the image contains a mask with a hole, when the image contains a disjoined mask consisting of multiple elements. Use pytest.mark.parametrize as demonstrated in this test.
  • code
  • documentation
  • unit tests
• Integration of rle_to_mask method into sv.DetectionDataset.from_coco logic.
  • code
  • documentation updates
• Integration of mask_to_rle method into sv.DetectionDataset.as_coco logic.
  • code
  • documentation updates

[emSko]

@SkalskiP what do you mean by the documentation? Only docstrings or are there additional files? Where to find them?

[SkalskiP]

Hi @emSko 👋🏻 Thank you very much for dedicating time to this task. In the meantime, I did some light research and found these two videos on YouTube:

• https://www.youtube.com/watch?v=tScp754b8iU
• https://www.youtube.com/watch?v=2zCZXZOoQJQ

I compared the results of our implementation of encoding and decoding as well as the execution time. The outcomes are identical; however, the execution time, especially for encoding, is significantly slower. Could we utilize similar tricks to accelerate the encoding process? Here's the Colab I used.

https://colab.research.google.com/drive/1Yl3w0B12htaso_VxZ-S2LvvxftiRk-ta?usp=sharing

[SkalskiP]

@SkalskiP what do you mean by the documentation? Only docstrings or are there additional files? Where to find them?

I would like the mask_to_rle and rle_to_mask functions to be included here. To make this happen, you need to add the appropriate changes to this file.

[LinasKo]

Hi @emSko 👋

Do a git pull - I've added a line that fixes the segfault.

I couldn't reproduce it as well, but found this issue

We'll eventually update the opencv to >=4.6.0.0 which allegedly resolves it, but for now this should be sufficient.

[emSko]

Hi @LinasKo, thank you for your help.

@SkalskiP I think the PR is ready for review. I don't think that from_coco requires any documentation updates.

[SkalskiP]

Hi @emSko 👋🏻 At the outset, I apologize for the late reply - I was traveling at the end of the week and was not on GitHub. @LinasKo, thanks for your support! 🙏🏻

@emSko you did an excellent job here! We are already on the last straight. I left some final comments.

[SkalskiP]

Hi @emSko 👋🏻

I just tested your code in Google Colab. Everything seems to be working fine. I also made a few minor changes to the docs, including adding some visualizations. Everything looks great. Many thanks for this work. This is a massive upgrade for COCO dataset support. Working with you was a pleasure. Thank you! 🙏🏻