Yolov7 Tiny setting confidence Thres about yolov7_openvino_cpp-python HOT 4 CLOSED

Hi @bbartling does this issue happen only on V7 or both on V7 and Tiny ? I can try it later with my webcam.
For model quantization part, one thing I must mention is the postprocess part of YOLOv7 model network in OpenVINO notebook example is a little different from the model in this repo. I believe the 3 outputs of model in notebook have been consolidated as one, so you have to modify your script accordingly. and you can also use Netron to find their difference.

from yolov7_openvino_cpp-python.

Hi Ethan, so the Yolov7 (non Tiny) works really well at classifying people from any angle....by far exceeds any model I've ever tested which is quite a few from the pyimagesearch caffee models as well as all the pretrained openVINO models for people/pedestrians. Problem is it just runs slow especially on CPU but for a conference room application 1-2 FPS is just fine as people dont move very fast when they are sitting : )

It seems like the Tiny Yolov7 reminds of about the same performance as the pretrained openVINO models for people/pedestrians, like some angles of the person the model performs very poor and when charted on a Grafana time series chart over time the graph is very jumpy....like it cannot be used to control purposes.

Ill experiment with the Yolov7 quantization, getting the code to run real curious to put this thru its paces.

from yolov7_openvino_cpp-python.

Any chance would you be able to elaborate where in the Yolo v7 optimization notebook on how to:

I believe the 3 outputs of model in notebook have been consolidated as one, so you have to modify your script accordingly. and you can also use Netron to find their difference.

On Ubuntu I can get down to here and eventually the kernel just dies....

mp, mr, map50, map, maps, num_images, labels = test(data=data, model=compiled_model, dataloader=dataloader, names=NAMES)
# Print results
s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Labels', 'Precision', 'Recall', '[email protected]', '[email protected]:.95')
print(s)
pf = '%20s' + '%12i' * 2 + '%12.3g' * 4  # print format
print(pf % ('all', num_images, labels, mp, mr, map50, map))

Any tips to try if you have the time greatly appreciated : )

from yolov7_openvino_cpp-python.

hi @bbartling , After you get a quantized model according to the guild in yolov7 notebook, you can switch the post processing part of this repo to which in yolov7 notebook sample, or just use the inference code in notebook sample, not this repo.

For the issue of kernel down, you can ask for help in yolov7 notebook repo directly.

from yolov7_openvino_cpp-python.