NMS-非极大值抑制

• 非极大值抑制算法（Non-maximum suppression, NMS）的本质是搜索局部极大值，抑制非极大值元素。
• 目标检测中应用NMS算法的主要目的是消除多余（交叉重复）的窗口，找到最佳物体检测位置。

目标检测可能会出现的一个问题是，模型会对同一目标做出多次预测，得到多个BBox。NMS旨在保留最接近真实包围盒(ground truth)的那一个预测结果，而抑制其他的预测结果。NMS的做法是，首先对每个类别，NMS先统计每个预测结果输出的属于该类别概率，并将预测结果按该概率由高至低排序。其次，NMS认为对应概率很小的预测结果并没有找到目标，所以将其抑制。然后，NMS在剩余的预测结果中，找到对应概率最大的预测结果，将其输出，并抑制和该包围盒有很大重叠(比如IoU大于0.3)的其他包围盒。重复上一步，直到所有的预测结果均被处理。

Python代码如下:

def non_max_suppression(boxes, max_bbox_overlap, scores=None):
    """Suppress overlapping detections.

    Original code from [1]_ has been adapted to include confidence score.

    .. [1] http://www.pyimagesearch.com/2015/02/16/
           faster-non-maximum-suppression-python/

    Examples
    --------

        >>> boxes = [d.roi for d in detections]
        >>> scores = [d.confidence for d in detections]
        >>> indices = non_max_suppression(boxes, max_bbox_overlap, scores)
        >>> detections = [detections[i] for i in indices]

    Parameters
    ----------
    boxes : ndarray
        Array of ROIs (x, y, width, height).
    max_bbox_overlap : float
        ROIs that overlap more than this values are suppressed.
    scores : Optional[array_like]
        Detector confidence score.

    Returns
    -------
    List[int]
        Returns indices of detections that have survived non-maxima suppression.

    """
    if len(boxes) == 0:
        return []

    boxes = boxes.astype(np.float)
    pick = []

    x1 = boxes[:, 0]
    y1 = boxes[:, 1]
    x2 = boxes[:, 2] + boxes[:, 0]
    y2 = boxes[:, 3] + boxes[:, 1]

    area = (x2 - x1 + 1) * (y2 - y1 + 1)
    if scores is not None:
        idxs = np.argsort(scores)
    else:
        idxs = np.argsort(y2)

    while len(idxs) > 0:
        last = len(idxs) - 1
        i = idxs[last]
        pick.append(i)

        xx1 = np.maximum(x1[i], x1[idxs[:last]])
        yy1 = np.maximum(y1[i], y1[idxs[:last]])
        xx2 = np.minimum(x2[i], x2[idxs[:last]])
        yy2 = np.minimum(y2[i], y2[idxs[:last]])

        w = np.maximum(0, xx2 - xx1 + 1)
        h = np.maximum(0, yy2 - yy1 + 1)

        overlap = (w * h) / area[idxs[:last]]

        idxs = np.delete(
            idxs, np.concatenate(
                ([last], np.where(overlap > max_bbox_overlap)[0])))

    return pick