def build_rpn_targets(anchors, gt_class_ids, gt_boxes, config):
    """Given the anchors and GT boxes, compute overlaps and identify positive
    anchors and deltas to refine them to match their corresponding GT boxes.

    anchors: [num_anchors, (y1, x1, y2, x2)]
    gt_class_ids: [num_gt_boxes] Integer class IDs.
    gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)]

    Returns:
    rpn_match: [N] (int32) matches between anchors and GT boxes.
               1 = positive anchor, -1 = negative anchor, 0 = neutral
    rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas.
    """
    # RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
    rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
    # RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))]
    rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4))

    # Handle COCO crowds
    # A crowd box in COCO is a bounding box around several instances. Exclude
    # them from training. A crowd box is given a negative class ID.
    crowd_ix = np.where(gt_class_ids < 0)[0]
    if crowd_ix.shape[0] > 0:
        # Filter out crowds from ground truth class IDs and boxes
        non_crowd_ix = np.where(gt_class_ids > 0)[0]
        crowd_boxes = gt_boxes[crowd_ix]
        gt_class_ids = gt_class_ids[non_crowd_ix]
        gt_boxes = gt_boxes[non_crowd_ix]
        # Compute overlaps with crowd boxes [anchors, crowds]
        crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes)
        crowd_iou_max = np.amax(crowd_overlaps, axis=1)
        no_crowd_bool = (crowd_iou_max < 0.001)
    else:
        # All anchors don't intersect a crowd
        no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool)

    # Compute overlaps [num_anchors, num_gt_boxes]
    overlaps = utils.compute_overlaps(anchors, gt_boxes)

    # Match anchors to GT Boxes
    # If an anchor overlaps a GT box with IoU >= 0.7 then it's positive.
    # If an anchor overlaps a GT box with IoU < 0.3 then it's negative.
    # Neutral anchors are those that don't match the conditions above,
    # and they don't influence the loss function.
    # However, don't keep any GT box unmatched (rare, but happens). Instead,
    # match it to the closest anchor (even if its max IoU is < 0.3).
    #
    # 1. Set negative anchors first. They get overwritten below if a GT box is
    # matched to them. Skip boxes in crowd areas.
    anchor_iou_argmax = np.argmax(overlaps, axis=1)
    anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
    rpn_match[(anchor_iou_max < 0.3) & (no_crowd_bool)] = -1
    # 2. Set an anchor for each GT box (regardless of IoU value).
    # TODO: If multiple anchors have the same IoU match all of them
    gt_iou_argmax = np.argmax(overlaps, axis=0)
    rpn_match[gt_iou_argmax] = 1
    # 3. Set anchors with high overlap as positive.
    rpn_match[anchor_iou_max >= 0.7] = 1

    # Subsample to balance positive and negative anchors
    # Don't let positives be more than half the anchors
    pos_ids = np.where(rpn_match == 1)[0]
    extra = len(pos_ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2)
    if extra > 0:
        # Reset the extra ones to neutral
        pos_ids = np.random.choice(pos_ids, extra, replace=False)
        rpn_match[pos_ids] = 0
    # Same for negative proposals
    neg_ids = np.where(rpn_match == -1)[0]
    extra = len(neg_ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE -
                        np.sum(rpn_match == 1))
    if extra > 0:
        # Rest the extra ones to neutral
        neg_ids = np.random.choice(neg_ids, extra, replace=False)
        rpn_match[neg_ids] = 0

    # For positive anchors, compute shift and scale needed to transform them
    # to match the corresponding GT boxes.
    ids = np.where(rpn_match == 1)[0]
    ix = 0
    # TODO: use box_refinment() rather than duplicating the code here
    for idx, anchor in zip(ids, anchors[ids]):
        # Closest gt box (it might have IoU < 0.7)
        gt = gt_boxes[anchor_iou_argmax[idx]]

        # Convert coordinates to center plus width/height.
        # GT Box
        gt_h = gt[2] - gt[0]
        gt_w = gt[3] - gt[1]
        gt_center_y = gt[0] + 0.5 * gt_h
        gt_center_x = gt[1] + 0.5 * gt_w
        # Anchor
        a_h = anchor[2] - anchor[0]
        a_w = anchor[3] - anchor[1]
        a_center_y = anchor[0] + 0.5 * a_h
        a_center_x = anchor[1] + 0.5 * a_w

        # Compute the bbox refinement that the RPN should predict.
        rpn_bbox[ix] = [
            (gt_center_y - a_center_y) / a_h,
            (gt_center_x - a_center_x) / a_w,
            np.log(gt_h / a_h),
            np.log(gt_w / a_w),
        ]
        ix += 1

    # Normalize
    rpn_bbox[ix] /= config.BBOX_STD_DEV
    return rpn_match, rpn_bbox