Anchor Matching via IoU Assignment

MediumComputer Vision

Implement match_anchors(anchors, gt_boxes, pos_threshold=0.5, neg_threshold=0.4), the anchor-to-ground-truth matcher used by one-stage detectors.

anchors: (N, 4) and gt_boxes: (M, 4), xyxy format [x1, y1, x2, y2]; lists or NumPy arrays.

Procedure:

Pairwise IoU matrix (N, M): $\mathrm{IoU}(a,g)=|a\cap g|/|a\cup g|$ ; an empty intersection or a zero-area union gives 0.
For each anchor take its max IoU over the GTs and the argmax GT.
Label: IoU $\ge$ pos_threshold → 1 with matched_gt = argmax GT; IoU $<$ neg_threshold → 0; otherwise ignore, -1. Non-positive anchors get matched_gt = -1.
Force a match for every GT: for each GT j in order, its highest-IoU anchor becomes 1 with matched_gt = j, even below pos_threshold; a later j overrides an earlier one on the same anchor.

Edge cases: M = 0 → all anchors 0 with matched_gt = -1; N = 0 → two empty int arrays of shape (0,).

Return (labels, matched_gt): integer np.ndarrays of shape (N,); labels take values in {1, 0, -1}.

Example:

Input:

anchors = [[0,0,10,10], [5,5,15,15], [20,20,30,30], [0,0,2,2]]
gt_boxes = [[0,0,10,10], [18,18,28,28]]
pos_threshold = 0.5, neg_threshold = 0.4

Output:

labels = [1, 0, 1, 0]
matched_gt = [0, -1, 1, -1]

Reasoning:

IoU row-maxima are ≈ [1.00, 0.14, 0.47, 0.04]. Thresholding alone marks anchor 0 positive (GT 0), anchors 1 and 3 negative, and anchor 2 ignore (0.4 ≤ 0.47 < 0.5). Forced GT matching then promotes anchor 2 to positive for GT 1 (it is GT 1's best anchor), yielding labels [1, 0, 1, 0] and matched_gt [0, -1, 1, -1].

Learn About topic

Contributors:

Deep-ML

Contribute