Different types of IoU

[kennyFF92]

🚀 Feature

Implementation of 3 other types of IoU metric in torchvision.ops.boxes:

1. Generalized IoU
2. Distance IoU
3. Complete IoU

Motivation

Since these metrics can lead to better results in some cases (object detection, etc.) it could be useful to have them natively implemented.

Pitch

Users who need IoU computation can simply choose from 4 different implementation using the same inputs (e.g. two lists of boxes).

Additional context

Mathematical details can be found here:

1. Generalized IoU: Generalized Intersection over Union
2. Distance/Complete IoU: Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression

[fmassa]

Hi,

This is a good request. I have seen benefits for Generalized IoU in the past, although for us Distance / Complete IoU didn't show any benefits.

From a number of citations perspective, I would say that we could currently add Generalized IoU to torchvision, and wait a bit more on Distance / Complete IoU to see if more people find it helpful.

[fmassa]

For information, here is an implementation for Generalized IoU that fits the same API as the IoU from torchvision

[kennyFF92]

For information, here is an implementation for Generalized IoU that fits the same API as the IoU from torchvision

I already had an implementation and now I aligned it to the one you mentioned. I'd like to contribute with a pull request but I can't find a guide for contributing and I don't want to create too many problems

[fmassa]

For contributing, the standard guidelines apply:

• fork the repo and create a branch for the feature
• add tests and documentation
• send the pull request and make sure that all tests pass (including lint)

[oke-aditya]

Hello, @fmassa can I contribute to this issue by creating a PR?
I am new to contributing to torchvision but highly interested.
I will try my best to add this but please do help me.