Extension to object detection about cleanlab

[jwmueller]

We will be providing an official tutorial on how to best use cleanlab for object detection in the near future, so stay tuned!

[jwmueller]

Update status on the tutorial is we're now wrapping up project introducing the proper usage of cleanlab for semantic segmentation, and then will extend the same strategy to object detection after that. Ensuring we are contributing the best algorithms into this repo rather than just some quick hacks requires us to do extensive benchmarking on many datasets, so thanks for your patience!

[sisrfeng]

I have read this :

If your regression targets were probabilities (between 0 and 1), it's possible it could handle up to 3 decimals of target granularity for regression.
But I have not run these experiments.
Throughout our work, we use the term "label" to refer to a "class label",
reserving "targets" for regression target values.
And we only deal with noisy "labels".

let d=1e-3
0-1 as 1k classes can be like:
0, 1d, 2d, 999d

object detection models like YOLO:
We reframe object detection as a single regression problem, straight from image pixels to bounding box coordinates and class probabilities.

I think object detection task may be reduced to what cleanlab can handle, but I haven't got the complete idea yet.
I will try and share my experience.

Thx!

[jwmueller]

@ibayer We are aiming to have one by early July!
If you want to DM me details about your use-case, I can show you how you can currently apply the current cleanlab v2.0.0 for object detection. You can message me in our Slack channel:

https://join.slack.com/t/cleanlab-community/shared_invite/zt-17lszn4hv-gg2FhZPXYfljq_l01uo92g

[AllenDun]

also expected feature, it will be very useful.

[cgnorthcutt]

Thank you for our suggestions.
@AllenDun @DuinoDu Expected futures we plan to release are now available here: https://github.com/cgnorthcutt/cleanlab/blob/master/cleanlab/version.py

A link has been added to the README.

[sisrfeng]

Hi！
So far there is no github repo which can find the wrong labels in object detection datasets, right?
I can't find any

[cgnorthcutt]

Hi, if the problem can be represented as probabilities over classes and noisy labels for each example, then cleanlab can be used. Can you think of how to reduce your task to this? Cleanlab should be able to work for this. Let me know if you run into an issue with regards to cleanlab.

[ibayer]

@jwmueller
I just stumbled across this issue from a google search for cleanlab object detection.

We will be providing an official tutorial

Have you been able to publish the tutorial already? I haven't used cleanlab before and this could
be great entry point for me and others.

[ibayer]

@jwmueller

How about re-opening the issue till the

an official tutorial on how to best use cleanlab for object detection

has been published?

If you want to DM me details about your use-case,

Sure, I'm looking for an initial POC that shows how well cleanlab works on some popular object detection dataset (coco?) and how much effort it is to set this up.

If it works well on coco, I'm positive it will transfer well to user use-cases, I can share at that point.

[jwmueller]

@ibayer A blogpost exploring the use of cleanlab for object detection for coco would be amazing! I'm happy to give some pointers if you can ping me in our slack channel where it's easier to have interactive discussion:
https://cleanlab.ai/slack

[jwmueller]

will be closed by: #676

[alexdesion]

Hello.

What's the current status of this addition? I am following the roadmap a long time ago. I am pretty sure that this functionality will be interesting for a lot of people.

Thank you for your work!

[jwmueller]

@alexdesion There is a PR adding support for label error detection in object detection here:
#676

Feel free to try it out and let us know how it works for you! We will be merging it soon, just polishing various things but the algorithm there is all finalized