yolo通过聚类设计anchor

by 目标检测原理qq群：703346870
实际工程中需要设计新的anchor来满足自身数据集。
本代码目的使用kmeans聚类出anchor的长宽

#-*-coding:utf-8-*-
'''
by 目标检测原理qq群：703346870
实际工程中需要设计新的anchor来满足自身数据集。
本代码目的使用kmeans聚类出anchor的长宽
参考
[1] 博客https://blog.csdn.net/cgt19910923/article/details/82154401
[2] git https://github.com/lars76/kmeans-anchor-boxes
'''
import glob
import xml.etree.ElementTree as ET
import numpy as np
import pdb
 
def iou(box, clusters):
    """
    Calculates the Intersection over Union (IoU) between a box and k clusters.
    :param box: tuple or array, shifted to the origin (i. e. width and height)
    :param clusters: numpy array of shape (k, 2) where k is the number of clusters
    :return: numpy array of shape (k, 0) where k is the number of clusters
    """
    x = np.minimum(clusters[:, 0], box[0])
    y = np.minimum(clusters[:, 1], box[1])
    if np.count_nonzero(x == 0) > 0 or np.count_nonzero(y == 0) > 0:
        raise ValueError("Box has no area")

    intersection = x * y
    box_area = box[0] * box[1]
    cluster_area = clusters[:, 0] * clusters[:, 1]

    iou_ = intersection / (box_area + cluster_area - intersection)

    return iou_

def avg_iou(boxes, clusters):
    """
    Calculates the average Intersection over Union (IoU) between a numpy array of boxes and k clusters.
    :param boxes: numpy array of shape (r, 2), where r is the number of rows
    :param clusters: numpy array of shape (k, 2) where k is the number of clusters
    :return: average IoU as a single float
    """
    return np.mean([np.max(iou(boxes[i], clusters)) for i in range(boxes.shape[0])])

def translate_boxes(boxes):
    """
    Translates all the boxes to the origin.
    :param boxes: numpy array of shape (r, 4)
    :return: numpy array of shape (r, 2)
    """
    new_boxes = boxes.copy()
    for row in range(new_boxes.shape[0]):
        new_boxes[row][2] = np.abs(new_boxes[row][2] - new_boxes[row][0])
        new_boxes[row][3] = np.abs(new_boxes[row][3] - new_boxes[row][1])
    return np.delete(new_boxes, [0, 1], axis=1)

def kmeans(boxes, k, dist=np.median):
    """
    Calculates k-means clustering with the Intersection over Union (IoU) metric.
    :param boxes: numpy array of shape (r, 2), where r is the number of rows
    :param k: number of clusters
    :param dist: distance function
    :return: numpy array of shape (k, 2)
    """
    rows = boxes.shape[0]

    distances = np.empty((rows, k))
    last_clusters = np.zeros((rows,))

    np.random.seed()

    # the Forgy method will fail if the whole array contains the same rows
    clusters = boxes[np.random.choice(rows, k, replace=False)]

    while True:
        for row in range(rows):
            distances[row] = 1 - iou(boxes[row], clusters)

        nearest_clusters = np.argmin(distances, axis=1)

        if (last_clusters == nearest_clusters).all():
            break

        for cluster in range(k):
            clusters[cluster] = dist(boxes[nearest_clusters == cluster], axis=0)

        last_clusters = nearest_clusters

    return clusters
 
 
def load_dataset(path):
    dataset = []
    for xml_file in glob.glob("{}/*xml".format(path)):
        tree = ET.parse(xml_file)        
        height = int(tree.findtext("./size/height"))
        width = int(tree.findtext("./size/width"))
        
        for obj in tree.iter("object"):
            xmin = int(obj.findtext("bndbox/xmin")) / width
            ymin = int(obj.findtext("bndbox/ymin")) / height
            xmax = int(obj.findtext("bndbox/xmax")) / width
            ymax = int(obj.findtext("bndbox/ymax")) / height
            xmin = np.float64(xmin)
            ymin = np.float64(ymin)
            xmax = np.float64(xmax)
            ymax = np.float64(ymax)
            if xmax == xmin or ymax == ymin:
                print(xml_file)
            dataset.append([xmax - xmin, ymax - ymin])
    return np.array(dataset)
 
if __name__ == '__main__':
    ANNOTATIONS_PATH = "../detection/val" #xml文件所在文件夹
    CLUSTERS = 3#9 #聚类数量，anchor数量
    INPUTDIM = 416 #输入网络大小
 
    data = load_dataset(ANNOTATIONS_PATH)
    out = kmeans(data, k=CLUSTERS)
    print('Boxes:')
    print(np.array(out)*INPUTDIM)    
    print("Accuracy: {:.2f}%".format(avg_iou(data, out) * 100))        
    #print(out)
    #ratios = np.around(out[:, 0] / out[:, 1], decimals=2).tolist()
    #print("Ratios:\n {}".format(sorted(ratios)))

 

参考
[1] 博客https://blog.csdn.net/cgt19910923/article/details/82154401
[2] git https://github.com/lars76/kmeans-anchor-boxes