MatrixNMS-8 spec.

[ceciliapeng2011]

Details:

• Specification of a new operator called MatrixNMS.
• This Op functionally equivalent to NonMaxSuppression-5, with an non max suppression algorithm called matrix_nms, and perform more post-processing phases, and lay out the detection outputs in the way of PaddlePaddle detection.

Tickets:

• 55968

[lazarevevgeny]

@openvinotoolkit/openvino-docs-maintainers , please, review this PR.

[ceciliapeng2011]

@zhangYiIntel welcome to review this spec, especially the operation algorithm description.

[zhangYiIntel]

PaddlePaddle has an attribute called normalized to specify whether the bounding boxes are normalized by the image width/height. Do we need to add this attribute ?

[ceciliapeng2011]

Don't have to be in OV Op. We will handle it in our PDPD->OV conversion.

[lazarevevgeny]

How do you know the image size if you want to handle it in PDPD->OV conversion? So I suggest to add this attribute.

[zhangYiIntel]

During op conversion, it gets the attribute value True/False, but during calculation, they have different IOU calculation methods depends on the normalized value. how can we handle this during the Ops conversion ?

[ceciliapeng2011]

This "normalize" attribute is PDPD specific, no need to use image size. It simply add a const offset to the original bbox. And as discussed with ZhangYi, we can handle this with a decomposition.

Please refer to Paddle2ONNX, which solve this by adding 2 Add operation before NMS.
https://github.com/PaddlePaddle/Paddle2ONNX/blob/e95695af98f1cf417d3ba2c0e81abf7b59948ab6/paddle2onnx/op_mapper/detection/multiclass_nms.py#L106

[ceciliapeng2011]

@lazarevevgeny I am so sorry I have to reopen this discussion. During test I found it cannot be solved in decomposition. Extra substract operation is required after NMS. It looks awkward. What's more, the extra substracts won't success as they are dynamic shape input.