Vectorize `compute_iou_matrix` with NumPy

[dpascualhe]

In our detection_metrics module, compute_iou_matrix currently uses a Python double for loop to fill a matrix of IoU values, which is slow for large datasets.

# current (slow)
for i, pb in enumerate(pred_boxes):
    for j, gb in enumerate(gt_boxes):
        iou_matrix[i, j] = compute_iou(pb, gb)

Deliverables: rewrite said function using fully vectorized NumPy broadcasting to compute all pairwise IoUs at once and make sure that the new implementation still yields the exact same results

[Arkadeep455]

Hi there,
I have started working on this issue.

[dpascualhe]

@Arkadeep455 , before starting to work in this issue. How do you plan to tackle it down?

[Arkadeep455]

So, my first approach even before I start any coding is to analyze the code and identify the design pattern which in this case is I think is a creation pattern. Once, I have fully analyzed the code and understood the problem statement, I write the code in different languages to check it's execution time. Since, this problem statement is all about calculations, I have chosen C language as it will speeds up calculations and to achieve better optimization.

After all this I went after the main problem statement. My first approach was parallel processing after few iterations I started all over again with NumPy broadcasting after some research.

Also, I tried to run only that loop in C so that the code remains same and the loop gets executed in hardware level and it worked out well. I am planning to include that part inside comments.

Shall I make a pull request now?