bbox reprojection method “corner”

[柴不燃]

if method in ('balanced', 'side_midpoints'):
old_side_midpoints = center + np.stack([-dx, -dy, dx, dy])
new_side_midpoints = cameralib.reproject_image_points(
old_side_midpoints, old_camera, new_camera)
new_midpoint_box = boxlib.bb_of_points(new_side_midpoints)

if method in ('balanced', 'corners'):
old_corners = center + np.stack([-dx - dy, dx - dy, dx + dy, -dx + dy])
new_corners = cameralib.reproject_image_points(
old_corners, old_camera, new_camera)
new_corner_box = boxlib.bb_of_points(new_corners)
center + np.stack([-dx - dy, dx - dy, dx + dy, -dx + dy])； i can't guess what you think？
center + np.stack([-dx - dy, dx - dy, dx + dy, -dx + dy]) ==>center + np.stack([-dx, -dy, dx, dy]) + np.stack([-dy, dx, dy, -dx])
is this an error? Is this what you really think？
center + np.stack([-dx, -dy, dx, dy]) + np.stack([-dy, -dx, dy, dx])==》 center + np.stack([-dx - dy, -dx - dy, dx + dy, dx + dy])

[István Sárándi]

Perhaps the variables are named a bit confusingly.

Here dx and dy are both 2D vectors:

    dx = np.array([old_box[2] / 2, 0])
    dy = np.array([0, old_box[3] / 2])

So center + np.stack([-dx - dy, dx - dy, dx + dy, -dx + dy]) gives the corner points.

Big picture about this function: the idea is to take an axis-aligned 2D bounding box and reproject it into another axis-aligned bounding box for a new camera view. For this, two options are to either transform the four corner points and take their bounding box, or to transform the midpoints of the sides and take the bbox of those transformed points. Then the 'balanced' option takes the average of those two results.