RPi Zero 2W, Coral TPU USB Acc, Hand Pose Tracking - Which model to use?

[awkyu]

Issue Type

Others

OS

Ubuntu

OS architecture

Other

Programming Language

Python

Framework

TensorFlowLite

Model name and Weights/Checkpoints URL

033_Hand_Detection_and_Tracking

https://github.com/PINTO0309/PINTO_model_zoo/tree/main/033_Hand_Detection_and_Tracking

Description

Hello!

I am working on an RPi Zero 2W with a Coral TPU USB Accelerator and was looking to implement Hand Tracking. Specifically, I wanted to get handpose from a camera output also connected to the rpi. I was a little confused on which model in 033_Hand_Detection_and_Tracking to use that would be best for this application. Could you provide more information on which model would be best to use?

Thanks!

Relevant Log Output

No response

URL or source code for simple inference testing code

No response

[PINTO0309]

Any of these are fine.

[awkyu]

Could you possibly explain some of the tensor outputs from all the models? I'm looking at specifically the integer quantization and see these size outputs:

1. hand_landmark_256_integer_quant.tflite - 1x42 tensor (is this 2d hand landmarks?) and 1x1 tensor
2. hand_landmark_new_256x256_integer_quant.tflite - 1x1x1x1 tensor, 1x1x1x1 tensor, 1x63 tensor (which is hand landmarks (I assume world)
3. hand_landmark_3d_256_integer_quant.tflite - 1x63 tensor (which is hand landmarks (I assume world) and 1x1 tensor
4. palm_detection_builtin_256_integer_quant.tflite - 1x2944x1 tensor (which I assume is confidence values for anchors) and 1x2944x18 tensor (0-3 being bounding box, 4-17 being 7 keypoints x and y positions)

Thanks, I'm pretty new to this, so any help would be appreciated! @PINTO0309