Cpp on yolov8 with hailo8l for pose_estimtion

General
1

Hi, I am trying to obtain keypoints on a person.jpg image by running the suggested script on c++ (Hailo-Application-Code-Examples/runtime/cpp/pose_estimation/yolov8_pose at dd6ada9d0d10e8b75660b74ab56ba018165204c0 · hailo-ai/Hailo-Application-Code-Examples · GitHub) and executing ./build/aarch64/vstream_yolov8pose_example_cpp -hef=yolov8s_h8l.hef -input=person.jpg, but no keypoints seem to be found, while the bounding box has been detected and drawn correctly on the image. I am trying this via yolov8s_pose_h8l.hef / yolov8s_pose.hef model (I do not think that the problem is the model) but values are out of scale. Instead the equivalent script taken from the same repository on python works. Why? Do you have any suggestions? I think that c++ has some quantization problem. Here’s the output. The fact that dequatization comes from (val-zero_point)*scale and the relevant tensors are 2, 5, 8 have strange values seems to be the problem: [DEBUG EXAMPLE] Keypoints quantization parameters:

  • Tensore 0 | Scale: 0.0798196 | Zero Point: 84
  • Tensore 1 | Scale: 0.00392157 | Zero Point: 0
  • Tensore 2 | Scale: 0.000618816 | Zero Point: 17224
  • Tensore 3 | Scale: 0.0740629 | Zero Point: 72
  • Tensore 4 | Scale: 0.00392157 | Zero Point: 0
  • Tensore 5 | Scale: 0.000525168 | Zero Point: 19190
  • Tensore 6 | Scale: 0.075078 | Zero Point: 62
  • Tensore 7 | Scale: 0.00392157 | Zero Point: 0
  • Tensore 8 | Scale: 0.000423056 | Zero Point: 18261. This means that keypoints become negative after dequantization [DEBUG:PREDEQUANTIZED KPTS] Debug Dequantization Keypoint
    qp_scale_kpts: 0.000618816, qp_zp_kpts: 17224
    Keypoints RAW (Quantized)
    [0]: 222, 70, 128
    [1]: 70, 119, 67
    [DEBUG:--------DEQUANTIZED KPTS] Keypoints Dequantized
    [0]: -10.5019, -10.6127, -10.6232
    [1]: -10.6238, -10.5879, -10.6034. And after the centring they get these absurde values. Here’s some: Number di keypoints: 17
    Keypoint [0]: (-416.123, -359.213), Confidence: 2.43436e-05
    Keypoint [1]: (-423.926, -357.628), Confidence: 2.48304e-05
    Keypoint [2]: (-425.351, -359.173), Confidence: 2.5031e-05
2

Hi @Simone_Tortorella
It looks like there is some mismatch in expected output types vs actual output types. The Tensors 2, 5, and 8 seem to be quantized using UINT16 as opposed to UINT8 and hence you see higher values.

3

Hi @Simone_Tortorella
Here is another thread in which this was discussed and finally resolved as well: Custom python Post processing function in gstreamer performance issues - General - Hailo Community

4

Yes, I had to cast to unit16 the bounding box output, which was unit8, and extract keypoints as unit16. Thanks!!

1 Like