SOFT-NMS (二) （non maximum suppression，非极大值抑制）

import numpy as np

boxes = np.array([[200, 200, 400, 400], [220, 220, 420, 420], [200, 240, 400, 440], [240, 200, 440, 400], [1, 1, 2, 2]], dtype=np.float32)
boxscores = np.array([0.9, 0.8, 0.7, 0.6, 0.5], dtype=np.float32)  # 不放在里面是因为scores会发生变化

def softnms(boxes_x1y1x2y2, score_boxes, threshold_iou = 0.3, sigma=0.5, threshold_end=0.001, method=2):
    """
    softnms
    :param boxes_x1y1x2y2:   boexs 坐标矩阵 format [y1, x1, y2, x2]
    :param score_boxes:     每个 boxes 对应的分数
    :param threshold_iou:           其余boxes与一个最好score的box分别求解iou，大于score_boxes值表示与该box重叠很高的重复框去掉，小于该值的可能是其它object的框， 交叠门限
    :param sigma:        使用 gaussian 函数的方差
    :param threshold_end:      最后的分数门限,将一次一次求解最好的box对应的iou值保存在scores中，而后保留大于此阈值的scores对应boxes
    :param method:      使用的方法
    :return:            留下的 boxes 的 index
    """
    # indexes concatenate boxes with the last column
    N = boxes_x1y1x2y2.shape[0]   # 表示boxes的个数
    # the order of boxes coordinate is [y1,x1,y2,x2]
    y1 = boxes_x1y1x2y2[:, 0]
    x1 = boxes_x1y1x2y2[:, 1]
    y2 = boxes_x1y1x2y2[:, 2]
    x2 = boxes_x1y1x2y2[:, 3]
    scores = score_boxes
    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
    for i in range(N):
        # intermediate parameters for later parameters exchange
        tBD = boxes_x1y1x2y2[i, :].copy()    # 对应的第i个坐标
        tscore = scores[i].copy()  # 对应的第i个scores
        tarea = areas[i].copy()   # 对应的第i个面积
        pos = i + 1               # pos为跳过当前第i个的box的索引
        if i != N - 1:            # 判断i是否取最大了
            maxscore = np.max(scores[pos:], axis=0)   # 跳过当前第i个后的scores分数取最大的
            maxpos = np.argmax(scores[pos:], axis=0)  # 跳过当前第i个后的scores分数取最大的分数对应的索引
        else:
            maxscore = scores[-1]     # 如果没有（即只有一个），就取这个
            maxpos = 0                # 因为只有一个，即索引为0
        if tscore < maxscore:         # 第一个与
            # 每次把最高score的 往上拿
            boxes_x1y1x2y2[i, :] = boxes_x1y1x2y2[maxpos + i + 1, :]  # 行置换，score，area也一样
            boxes_x1y1x2y2[maxpos + i + 1, :] = tBD
            tBD = boxes_x1y1x2y2[i, :]
            scores[i] = scores[maxpos + i + 1]
            scores[maxpos + i + 1] = tscore
            tscore = scores[i]
            areas[i] = areas[maxpos + i + 1]
            areas[maxpos + i + 1] = tarea
            tarea = areas[i]
        # IoU calculate
        xx1 = np.maximum(boxes_x1y1x2y2[i, 1], boxes_x1y1x2y2[pos:, 1])
        yy1 = np.maximum(boxes_x1y1x2y2[i, 0], boxes_x1y1x2y2[pos:, 0])
        xx2 = np.minimum(boxes_x1y1x2y2[i, 3], boxes_x1y1x2y2[pos:, 3])
        yy2 = np.minimum(boxes_x1y1x2y2[i, 2], boxes_x1y1x2y2[pos:, 2])
        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        ovr = inter / (areas[i] + areas[pos:] - inter)
        # Three methods: 1.linear 2.gaussian 3.original NMS
        if method == 1:  # linear
            weight = np.ones(ovr.shape)
            weight[ovr > threshold_iou] = weight[ovr > threshold_iou] - ovr[ovr > threshold_iou]
        elif method == 2:  # gaussian
            weight = np.exp(-(ovr * ovr) / sigma)
        else:  # original NMS
            weight = np.ones(ovr.shape)
            weight[ovr > threshold_iou] = 0     # 表示大于Nt的iou值就和比较的box重合很多，因此大于这个阈值的iou为0，则相当于该score为0
        scores[pos:] = weight * scores[pos:]   # 给定scores的值，若不好就删除
    # select the boxes and keep the corresponding indexes
    boxes_new = boxes_x1y1x2y2[scores > threshold_end]   # 判断求解的scores是否满足阈值，若满足就保留
    return boxes_new
# boxes and scores

boees_slect = softnms(boxes, boxscores, method=3)
print(boees_slect)

原始NMS结果如下：

线性soft-nms:

高斯nms：