[TODO] Implementation of the `ArgMax` force-replacement option for `NMS` with associated class score refinement.

[PINTO0309]

Issue Type

Others

OS

Linux

onnx2tf version number

1.14.5

onnx version number

1.13.1

onnxruntime version number

1.15.1

onnxsim (onnx_simplifier) version number

0.4.33

tensorflow version number

1.13.0

Download URL for ONNX

• Naruto Detection with post-process yolox_nano_with_post.onnx.zip

Parameter Replacement JSON

N/A

Description

1. Personal
2. Implementation of the ArgMax force-replacement option for NMS with associated class score refinement.

TFLite's NMS does not support Multi-class Non-Maximum-Suppression. Thus, faithfully converting the onnx NonMaxSuppresion to the specification breaks it down into several very redundant NMSs, as shown in the figure below. Add an option to disable simultaneous NMS by multiple classes on the same bounding box and force conversion to output only one class, as this is undesirable from a mobile inference performance perspective.

• ONNX NMS
• TFLite NMS x 16