The code behind the picture

3 teaching cells · Python standard library · Read-only

1. 1. Prepare the candidate and the four adjustments

   The candidate is the search window over the watermelon, as top-left corner plus width and height in picture pixels (origin top-left, y downward). The four adjustments are the trained box head's output for this window, copied at full precision; the reference box is not an input to them.

   In [ ]:

   import json
   import math
   
   x, y, w, h = 48, 24, 80, 80  # candidate: search window 13
   t = [-0.2061745477183015, 0.16483249426655214, -0.34782754792939563, -0.396589869704336]  # tx, ty, tw, th from the box head

2. 2. Decode the predicted box, then measure IoU

   Move the centre by tx widths and ty heights, multiply the sides by e to the tw and th, and convert back to corners. Only then compare with the reference, supplied separately: intersection sides are clamped at zero, union = area A + area B − intersection, and an empty union gives 0. Round only the report.

   In [ ]:

   cx, cy = x + w / 2 + t[0] * w, y + h / 2 + t[1] * h
   pw, ph = w * math.exp(t[2]), h * math.exp(t[3])
   pred = [cx - pw / 2, cy - ph / 2, cx + pw / 2, cy + ph / 2]  # x1, y1, x2, y2
   ref = [49, 52, 103, 103]  # reference corners, supplied separately
   
   def iou(a, b):
       iw = max(0, min(a[2], b[2]) - max(a[0], b[0]))
       ih = max(0, min(a[3], b[3]) - max(a[1], b[1]))
       inter = iw * ih
       union = (a[2] - a[0]) * (a[3] - a[1]) + (b[2] - b[0]) * (b[3] - b[1]) - inter
       return inter, union, (inter / union if union > 0 else 0)
   
   inter, union, value = iou(pred, ref)
   result = {
       "predicted": [round(v, 4) for v in pred],
       "intersection": round(inter, 4),
       "union": round(union, 4),
       "iou": round(value, 4),
       "window_iou": round(iou([x, y, x + w, y + h], ref)[2], 4),
   }

3. 3. Inspect the result

   Print the result as JSON so both languages use the same output format.

   In [ ]:

   print(json.dumps(result))

{
  "predicted": [
    43.2572,
    50.2822,
    99.7549,
    104.091
  ],
  "intersection": 2588.498,
  "union": 3205.5721,
  "iou": 0.8075,
  "window_iou": 0.4303
}