OpenVino-R5 and Yolo3: problem with inference result

[hamze60]

Hello ,

On OpenVino R5, I could successfully convert YoloV3 to IR model (using the Intel's default suggested json file for Yolo3). However, when I read the detection result in python, using OpenCV, (detection = net.forward()), detection format is changed and actually it is not obvious what is the result format.

- In original Yolo3-darnket, I used to read the detection output of 3 layers (yolo-82, yolo-94, yolo 106) in OpenCV, which have this format:
    output[ layer =0, or 1 or 2]'s size = NumberofDetections X 85. Where 85 = 4 (box coordinates) + 1? + 80 (probability of each class)
- However, after Yolo3 to IR conversion, there are totally 3 layers: conv_6, conv_14, conv_22. Detection outputs are in this format:
    output[ layer =0]'s size = 1 X 255 X 26 X 26.
    output[ layer =1]'s size = 1 X 255 X 52 X 52.
    output[ layer =2]'s size = 1 X 255 X 13 X 13.

I could not find anything on OpenVino regarding output format after conversion, and how to read inference result. Is there any suggestion?

[hamze60 · Accepted solution]

Thanks to PINTO, he developed a function, called "ParseYOLOV3Output", for this purpose.

PINTO's YOLOv3 project: https://github.com/PINTO0309/OpenVINO-YoloV3