Anchor Box Generation

The problem, from first principles

Predicting a box from scratch is hard for a network. Anchor-based detectors make it easier by laying down a dense grid of anchor boxes, fixed reference rectangles of several sizes and shapes at every position, and asking the network only to adjust each one. A feature map with fh x fw cells, whose stride is the number of image pixels per cell, gets anchors centred at each cell's centre. At every location there is one anchor per combination of scale (the square root of the box area) and aspect ratio (height divided by width). Each anchor keeps the area scale^2 while its shape changes with the ratio.

From theory to code

Implement generate_anchors.

Constraints

• generate_anchors(fh, fw, stride, scales, ratios) returns an array of shape (fh * fw * len(scales) * len(ratios), 4) of boxes (x1, y1, x2, y2).
• Cell (i, j) has centre cy = (i + 0.5) * stride, cx = (j + 0.5) * stride. For a scale s and ratio r = h / w, the anchor has w = s / sqrt(r) and h = s * sqrt(r), so w * h = s**2.
• Order: cells row-major (i outer, j inner), then for each cell scales outer and ratios inner.
• The box is (cx - w/2, cy - h/2, cx + w/2, cy + h/2).

Hints