Movenet and Posenet models not working on CPU

[suyash-narain]

Description

Hi,
I am trying to run posenet project on aarch64 device (raspberry pi 4) on CPU. I am using gstreamer. But on running the pose_camera.py script, no pose is detected. No line drawings can be seen as well. I have a USB camera connected as /dev/video0. Confidence scores for keypoints remains at 0;

When I switch over to using opencv, I get accurate pose detections with non-zero confidence scores.

My input command is:
$python3 pose_camera.py --model model.tflite --videosrc /dev/video0

my model.tflite is downloaded from https://raw.githubusercontent.com/google-coral/test_data/master/posenet/posenet_mobilenet_v1_075_353_481_16_quant_decoder.tflite

movenet model used: https://tfhub.dev/google/lite-model/movenet/singlepose/lightning/tflite/int8/4

On running, I see the video on the screen, with live fps and NPoses=1, but i don't see any actual pose detected using the drawn lines.
If i use it using opencv, i get pose outputs, but not using gstreamer. I even installed the posenet_decoder.so

How do I run this project on CPU using the CPU based models?
I get the same issue when using movenets on gstreamer too.

As a follow- up, how can I port it to run on movenets, using gstreamer? I can run movenet using opencv, but would like to use gstreamer without opencv.

Some help would be really appreciated. thanks

Click to expand!

Issue Type

Support

Operating System

Linux

Coral Device

Dev Board

Other Devices

Rapsberry Pi 4

Programming Language

Python 3.9

Relevant Log Output

No response