[Object detection pipeline] Lower threshold

[NielsRogge]

What does this PR do?

This PR proposes to lower the default threshold of the object detection pipeline. It is currently set to 0.9, which causes a lot of inference widgets to not show any detected objects.

[amyeroberts]

Thanks for updating!

Agreed, it's not BC but this seems like a better value

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SangbumChoi]

FYI, we usually use much more lower score threshold (e.g. 0.3) to see just tendency of the model. (+ Note that transformers has transformer architecture which have lower confidence level)

facebookresearch/detr#216

https://github.com/IDEA-Research/DINO/blob/main/inference_and_visualization.ipynb

[qubvel]

@NielsRogge should we merge it, or do you have something to add?

[NielsRogge]

I haven't tried the pipeline yet with various thresholds. Do you think we should lower it even more? It could result in too many detections. Wondering what the best default value is

[amyeroberts]

@NielsRogge We can always lower it further if necessary. I'd merge as-is and based on observations / feedback we can always update.

[qubvel]

With my experiments on cppe-5 dataset threshold=0.4..0.5 visually looks fine, otherwise, too many FP appear.

With a quick search in other frameworks, I found for some models:

• In Utlralitics - 0.25 / 0.3
• In Mmedection - 0.05 / 0.3 / 0.5

But in general, it depends on the model and dataset

[NielsRogge]

Ok thanks, will merge in that case.