Unable to understand the logic behind the `object_detection_demo_yolov3_async.py` demo

I noticed that the recent pull for the file object_detection_demo_yolov3_async.py did try to remove the otherwise confusing flattening of blob but I still can’t figure out how most of the results are parsed from this blob.

Is there some YOLO guide that you guys (@eizamaliev ) followed (except the official paper which is not so clear)?

Some example snippets that I don’t understand include

1. if is_proportional: in the line131. What was the purpose behind introducing this additional parameter to pass around the functions?

2. In the part

       # Process raw value
        x = (col + x) / params.side
        y = (row + y) / params.side

I can’t understand how the x,y (values) are added with row and col is right. I mean what is the logic behind it?

Please help me get a clear understanding of the YOLO decoding 😕

[eizamaliev]

Hi, @pra-dan. Thanks for your questions, I’ll try to answer them all.

1. Demo supports two resize type: with aspect ratio kept and without. Because initially was used simple resize, this behavior is saved as default. So is_proportional parameter tells to the function how it should rescale output boxes.

2. As you can know, YOLO generate local coordinates relative to the cell. To get coordinates relative to image, you should add cell coordinates. By dividing by cell count across corresponding direction, you normalize coordinates.

[pra-dan]

Yeah, we are good to go 👍🏽