deep-learning - 为什么我们使用 mAp 分数来评估深度学习中的目标检测器？

Question

在此 Tensorflow detection model zoo他们在不同的检测架构中提到了 COCO mAp 分数。他们还表示，mApp 分数越高，准确度越高。不明白的是这是如何计算的？它的最高分是多少？为什么这个 mAp 分数因数据集而异？

Answer 1

要了解 MAP(平均精度)，我会先从 AP(平均精度)开始。

Suppose we are searching for images of a flower and we provide our image retrieval system a sample picture of a rose (query), we do get back a bunch of ranked images (from most likely to least likely). Usually not all of them are correct. So we compute the precision at every correctly returned image, and then take an average.

示例:

If our returned result is 1, 0, 0, 1, 1, 1 where 1 is an image of a flower, while 0 not, then the precision at every correct point is:

Precision at each correct image = 1/1, 0, 0, 2/4, 3/5, 4/6
Summation of these precisions = 83/30
Average Precision = (Precision summation)/(total correct images) = 83/120

旁注:

本节提供了每个正确图像的精度计算背后的详细说明，以防您仍然对上述分数感到困惑。

出于说明目的，设 1, 0, 0, 1, 1, 1存储在一个数组中，所以 results[0] = 1 , results[1] = 0等等。

让 totalCorrectImages = 0, totalImagesSeen = 0, pointPrecision = 0 pointPrecision的公式是 totalCorrectImages / totalImagesSeen
在 results[0], totalCorrectImages = 1, totalImagesSeen = 1因此 pointPrecision = 1
自 results[1] != 1 , 我们忽略它但 totalImagesSeen = 2 && totalCorrectImages = 1
自 results[2] != 1 , totalImagesSeen = 3 && totalCorrectImages = 1
在 results[3], totalCorrectImages = 2, totalImagesSeen = 4因此 pointPrecision = 2/4
在 results[4], totalCorrectImages = 3, totalImagesSeen = 5因此 pointPrecision = 3/5
在 results[5], totalCorrectImages = 4, totalImagesSeen = 6因此 pointPrecision = 4/6

A simple way to interpret is to produce a combination of zeros and ones which will give the required AP. For example, an AP of 0.5 could have results like 0, 1, 0, 1, 0, 1, ... where every second image is correct, while an AP of 0.333 has 0, 0, 1, 0, 0, 1, 0, 0, 1, ... where every third image is correct.

For an AP of 0.1, every 10th image will be correct, and that is definitely a bad retrieval system. On the other hand, for an AP above 0.5, we will encounter more correct images than wrong in the top results which is definitely a good sign.

MAP 只是 AP 的扩展。您只需取一定数量查询的所有 AP 分数的平均值。以上对 AP 分数的解释也适用于 MAP。 MAP 范围从 0 到 100，越高越好。

AP 公式在 Wikipedia

MAP 公式在 Wikipedia

归功于此 blog

编辑我:

在对象检测方面应用相同的概念。在这种情况下，您将计算每个类(class)的 AP。这是由给定类别的精确召回曲线下的面积给出的。从这一点上，您可以找到他们达到 mAP 的平均值。

更多详情请引用 2012 Pascal VOC Dev Kit的3.4.1和4.4节。 .相关论文可查阅 here .