8-bit quantized models for the hand tracking

[freesokim]

Please make sure that this is a feature request.

System information (Please provide as much relevant information as possible)

• MediaPipe Solution (you are using): Hand Tracking (Palm detection, Hand Landmark)
• Programming language : C++/Python/Android Java

I think MediaPipe hand tracking is a highly efficient ML solution that runs in real-time.

Becasue the hand tracking is the 32-bit floating type ML model,
I have to use the GPU delegate supporting the 32-bit floating computation to accelerate the hand tracking on edge device.
(I know that there are only 32-bit floating point models for MediaPipe ML solutions except for object detection and face detection)

However, GPU is very expensive and the edge devices have limited resource and computational power.
So we need to use the NPU delegate for the edge devices, and almost of NPU (ex. Edge TPU and Hexagon NPU) require a fully 8-bit quantized model as the input.

Do you have a plan to release the fully 8-bit quantized models for hand tracking (palm and hand landmark)?

Thank you.

[Stephenfang51]

@freesokim hello sir
I have same demand with you, i remember you mentioned the
PINTO_model_zoo provided the quantized model, is it a fully int8 model ? thanks !

[sureshdagooglecom]

Hi @Stephenfang51 .
This is a feature and We have similar issue ,could you please check this issue.

[freesokim]

@Stephenfang51
Hi, Sorry for late reply.
I think PINTO_model_zoo provide the 8-bit quantized models, but I have not tried them yet.
My question is google have any plan to release the official 8-bit quantized models for the hand tracking.
Thanks.

[freesokim]

@Stephenfang51
Do you mean that google don't have any release plan for quantized model?
Thanks.

[dubrovin-sudo]

I also have this question. Is google have any plan to release the official 8-bit quantized models for the hand tracking?