目录
有原理步骤：
Soft-nms的实现过程可以分为几个步骤：
1. 输入预测框
输入神经网络预测输出的所有框，每个框有四个坐标和一个类别得分。
2. 对于每个框计算其权重
权重可以使用三种不同的函数：max、linear和Gaussian。
3. 重复以下步骤，直到不再有框被删除
（1）选出最高得分的框，令其权重为1，与第一个框进行交换。
（2）计算当前框与剩余框的重叠率。
（3）根据重叠率和选定的函数计算权重。
（4）根据权重更新每个框的得分。
（5）剔除得分小于设定阈值的框。
4. 输出筛选后的结果
代码：
def soft_nms(dets, sigma=0.5, Nt=0.3, threshold=0.001, method=1):
"""
PyTorch implementation of SoftNMS algorithm.
# Arguments
dets: detections, size[N,5], format[x1,y1,x2,y2,score]
sigma: variance of Gaussian function, scalar
Nt: threshold for box overlap, scalar
threshold: score threshold, scalar
method: 0=Max, 1=Linear, 2=Gaussian
# Returns
dets: detections after SoftNMS, size[K,5]
"""
# Indexes concatenate detection boxes with the score
N = dets.shape[0]
indexes = np.array([np.arange(N)])
dets = np.concatenate((dets, indexes.T), axis=1)
for i in range(N):
# intermediate parameters for later parameters exchange
si = dets[i, 4]
xi = dets[i, :4]
area_i = (xi[2] - xi[0] + 1) * (xi[3] - xi[1] + 1)
if method == 1: # Linear
weight = np.ones((N - i))
weight[0] = si
else: # Gaussian
# Compute Gaussian weight coefficients
xx = np.arange(i, N).astype(np.float32)
if method == 2:
sigma = 0.5
ii = np.ones((xx.shape[0], 1)) * i
# print(sigma)
# print((xx - ii).shape)
gauss = np.exp(-1.0 * ((xx - ii) ** 2) / (2 * sigma * sigma))
if method == 2:
weight = gauss
else:
weight = np.zeros((N - i))
weight[0] = 1.0
weight[1:] = gauss / np.sum(gauss)
# Sort boxes by score
idx = np.arange(i, N)
idx_max = np.argmax(dets[idx, 4])
idx_max += i
# Swap boxes and scores
dets[i, 4], dets[idx_max, 4] = dets[idx_max, 4], dets[i, 4]
dets[i, :4], dets[idx_max, :4] = dets[idx_max, :4], dets[i, :4]
dets[i, 5], dets[idx_max, 5] = dets[idx_max, 5], dets[i, 5]
# Compute overlap ratios
xx1 = np.maximum(dets[i, 0], dets[idx, 0])
yy1 = np.maximum(dets[i, 1], dets[idx, 1])
xx2 = np.minimum(dets[i, 2], dets[idx, 2])
yy2 = np.minimum(dets[i, 3], dets[idx, 3])
w = np.maximum(0.0, xx2 - xx1 + 1)
h = np.maximum(0.0, yy2 - yy1 + 1)
inter = w * h
# Update weights
if method == 0: # Max
weight[idx_max - i + 1:] = np.where(inter > Nt, 0.0, 1.0)
else: # Linear / Gaussian
weight_matrix = np.zeros((weight.shape[0], weight.shape[0]))
weight_matrix[0, :] = weight
weight_matrix[1:, :] = np.diag(weight[1:])
iou = inter / (area_i + dets[idx, 4] * (1 - inter))
weight[idx - i + 1] = np.matmul(weight_matrix, (1.0 - iou).reshape(-1, 1)).reshape(-1)
weight[idx_max - i + 1:] = np.where(iou > Nt, 0.0, weight[idx_max - i + 1:])
# Apply weight
dets[idx, 4] = dets[idx, 4] * weight
# Weigh small scores
suppress_small = np.where(dets[idx, 4] < threshold)[0]
dets[suppress_small + i, 4] = 0.0
# remove boxes lower than threshold
idx_keep = np.where(dets[:, 4] > 0)[0]
dets = dets[idx_keep]
return dets[:, :5]
加注释版：
soft_nms的优点
1，解决了物体挨得很近导致的漏检问题
2，需要增加的超参数很少，只增加了一个sigma，阈值nms本来也有，iou是算出来的
3，计算复杂度相对于nms没有增加，都是O(n^2)，n是bboxes的数量。
import numpy as np
# 定义一个nms函数
def soft_nms(dets, thresh=0.3, sigma=0.5): # score大于thresh的才能存留下来，当设定的thresh过低，存留下来的框就很多,所以要根据实际情况调参
'''
input：
dets: dets是(n,5)的ndarray，第0维度的每个元素代码一个框：[x1, y1, x2, y2, score]
thresh: float
sigma: flaot
output：
index
'''
x1 = dets[:, 0] # dets:(n,5) x1:(n,) dets是ndarray, x1是ndarray
y1 = dets[:, 1]
x2 = dets[:, 2]
y2 = dets[:, 3]
scores = dets[:, 4] # scores是ndarray
# 每一个候选框的面积
areas = (x2 - x1 + 1) * (y2 - y1 + 1) # areas:(n,)
# order是按照score降序排序的
order = scores.argsort()[::-1] # order:(n,) 降序下标 order是ndarray
keep = []
while order.size > 0:
i = order[0] # i 是当下分数最高的框的下标
# print(i)
keep.append(i)
# 计算当前概率最大矩形框与其他矩形框的相交框的坐标，会用到numpy的broadcast机制，得到的是向量
# 当order只有一个值的时候，order[1]会报错说index out of range，而order[1:]会是[]，不报错，[]也可以作为x1的索引，x1[[]]为[]
xx1 = np.maximum(x1[i], x1[order[1:]]) # xx1:(n-1,)的ndarray x1[i]:numpy_64浮点数一个，x1[order[1:]]是个ndarray，可以是空的ndarray，如果是空ndarray那么xx1为空ndarray，如果非空，那么x1[order[1:]]有多少个元素，xx1就是有多少个元素的ndarray。x1[]是不是ndarray看中括号内的是不是ndarray，看中括号内的是不是ndarray看中括号内的order[]的中括号内有没有冒号，有冒号的是ndarray，没有的是一个数。
yy1 = np.maximum(y1[i], y1[order[1:]])
xx2 = np.minimum(x2[i], x2[order[1:]])
yy2 = np.minimum(y2[i], y2[order[1:]])
# 计算相交框的面积,注意矩形框不相交时w或h算出来会是负数，用0代替
w = np.maximum(0.0, xx2 - xx1 + 1) # xx2-xx1是(n-1,)的ndarray，w是(n-1,)的ndarray, n会逐渐减小至1
# 当xx2和xx1是空的，那w是空的
h = np.maximum(0.0, yy2 - yy1 + 1)
inter = w * h # inter是(n,)的ndarray
# 当w和h是空的，inter是空的
# 计算重叠度IOU：重叠面积/（面积1+面积2-重叠面积）
eps = np.finfo(areas.dtype).eps # 除法考虑分母为0的情况，np.finfo(dtype).eps，np.finfo(dtype)是个类，它封装了机器极限浮点类型的数，比如eps，episilon的缩写，表示小正数。
ovr = inter / np.maximum(eps, areas[i] + areas[order[1:]] - inter) # n-1 #一旦（面积1+面积2-重叠面积）为0，就用eps进行替换
# 当inter为空，areas[i]无论inter空不空都是有值的，那么ovr也为空
# 更新分数
weight = np.exp(-ovr*ovr/sigma)
scores[order[1:]] *= weight
# 更新order
score_order = scores[order[1:]].argsort()[::-1] + 1
order = order[score_order]
keep_ids = np.where(scores[order]>thresh)[0]
order = order[keep_ids]
return keep
import numpy as np
import cv2
# 读入图片，录入原始人框（[x1, y1, x2, y2, score]）
image = cv2.imread('w.jpg')
boxes = np.array([[5, 52, 171, 270, 0.9999],
[13, 1, 179, 268, 0.9998],
[20, 7, 176, 262, 0.8998],
[7, 5, 169, 272, 0.9687],
[3, 43, 162, 256, 0.9786],
[10, 56, 167, 266, 0.8988]])
# 将框绘制在图像上
image_for_nms_box = image.copy()
for box in boxes:
x1, y1, x2, y2, score = int(box[0]), int(box[1]), int(box[2]), int(box[3]), box[4] # x:col y:row
image_for_nms_box = cv2.rectangle(image_for_nms_box, (x1, y1), (x2, y2), (0,255,0), 2)
cv2.imwrite("w_all.jpg", image_for_nms_box)
cv2.imshow('w_all', image_for_nms_box)
# 使用soft_nms对框进行筛选
keep = soft_nms(boxes)
soft_nms_boxs = boxes[keep]
# 将筛选过后的框绘制在图像上
image_for_nms_box = image.copy()
for box in soft_nms_boxs:
x1, y1, x2, y2, score = int(box[0]), int(box[1]), int(box[2]), int(box[3]), box[4]
image_for_nms_box = cv2.rectangle(image_for_nms_box, (x1, y1), (x2, y2), (0,255,0), 2)
# Syntax: cv2.imwrite(filename, image)
cv2.imwrite("w_soft_nms.jpg", image_for_nms_box)
cv2.imshow('w_soft_nms', image_for_nms_box)
cv2.waitKey()
cv2.destroyAllWindows()
724
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
