[Feature] Parse the model response

[angel-penchev]

Describe the feature you'd like
The YOLO neural network return spec is in shape (16, 16, 5, 12). The 16 x 16 spec represents image squares divisions (grid cells) with a side of 32 pixels (516/32 = 16). The 5 x 12 represents the 5 bounding box predictions a grid should propose. A bounding box prediction is with size 5 (x of box center, y of box center, w of the box, h of the box, probability that an object exists in this box) + 7 probabilities of a given class existing in the grid cell (classes listed here) = 12.

A much clearer explanation could be found here.

Now as the part you need to code. First you need to define two constants: MIN_SCORE = 0.5
and MIN_IOU = 0.45. Then you need to iterate over every grid cell (16 x 16 = 256 gird cells in total).
For each cell you iterate over the 5 bounding box predictions. If the 'probability that an object exists in this box' (the 5th element in the spec) is higher than MIN_SCORE and a given class probability is higher than MIN_IOU, then add one to the final dictionary for the given class. The dictionary should look something like this: { "bicycle": 0, "bus": 1, "car": 8, "horse": 0, "motorbike": 0, "person": 0, "train": 0 }.

Additional context