为何OpenCV的KNearest在K=3时返回结果为4而非5？

运行的Python代码 #

import cv2 as cv
import numpy as np

trainFeaturesData = [
    [2,2,2,2],
    [3,3,3,3],
    [4,4,4,4],
    [5,5,5,5],
    [6,6,6,6],
    [7,7,7,7],
]
trainFeatures = np.array(trainFeaturesData, dtype = np.float32)

trainLabelsData = [ 2, 3, 4, 5, 6, 7 ];
trainLabels = np.array(trainLabelsData, dtype = np.float32)

knn = cv.ml.KNearest_create()
knn.train(trainFeatures, cv.ml.ROW_SAMPLE, trainLabels)

testFeatureData = [[ 5, 5, 5, 5, ]]
testFeature = np.array(testFeatureData, dtype = np.float32)

for k in [1, 3]:
    print("------------ k = {} --------------\n".format(k));
    ret, results, neighbours ,dist = knn.findNearest(testFeature, k)
    print( "result:     {}\n".format(results) )
    print( "neighbours: {}\n".format(neighbours) )
    print( "distance:   {}\n".format(dist) )

运行输出 #

------------ k = 1 --------------

result:     [[5.]]

neighbours: [[5.]]

distance:   [[0.]]

------------ k = 3 --------------

result:     [[4.]] <= Why??

neighbours: [[5. 4. 6.]]

distance:   [[0. 4. 4.]]

原因分析 #

核心原因是OpenCV的KNN实现中，当出现投票平局时，会选择训练数据中最早出现的类别。

在这个案例里，K=3时，最近的三个邻居对应的标签是5、4、6，三者的投票数都是1票，形成平局。此时OpenCV会从这些平局的类别里，挑选在训练标签列表中最先出现的那个——训练标签里4的位置在5之前，所以最终返回结果为4。

你可以调整训练数据的顺序验证这个逻辑：比如把标签5对应的样本移到标签4的前面，重新运行代码，K=3时的结果就会变成5。

内容的提问来源于stack exchange，提问作者codeDom