Finding balls by colour

Pictures are arrays

To your program, img.rgb is just a (240, 320, 3) array: channels 0, 1 and 2 are red, green and blue. Comparing whole channels gives a mask, a boolean array marking the pixels that pass:

r, g, b = [rgb[..., i].astype(float) for i in range(3)]
red = (r > 50) & (g < 0.4 * r) & (b < 0.4 * r)

Convert to float first, because uint8 arithmetic wraps around at 255.

Use ratios, not levels

A ball's shadowed side is about three times darker. Shading scales all three channels by the same factor, so a rule like r > 150 loses the shadowed half while ratios such as $g/r$ don't change. Fully lit colours:

Surface	r	g	b
red ball	230	61	61
yellow ball	247	194	26
green ball	51	179	107
blue ball	64	107	230
table	146	111	80
orange arm joints	255	115	46

The table and the yellow ball have almost the same $g/r$. What sets yellow apart is its tiny $b/r$. (Real vision code often converts to HSV for the same reason: hue ignores brightness.)

Centroids

vs, us = np.nonzero(mask) returns the row and column of every selected pixel; the ball's centre is their mean, $(\bar u, \bar v)$. A ball covers about 100–180 pixels here, so ask for a minimum count (say 20) before you believe a mask. That's also how you report a missing ball.

Your task

Implement find_ball(rgb, color) for "red", "yellow", "green" and "blue". It returns (u, v), or None if the ball isn't visible. The starter shows every mask in the Images tab so you can see what your rules pick up.

The grader renders three new scenes. Each centroid must be within 3 px of the ball's true centre, and in one scene the blue ball is hidden under the table. For a bonus, it adds sensor noise: a few stray pixels far from the ball can drag a mean away, but a median barely moves.