探究平均准确度AP指标的缺陷及其影响

平均准确度AP是计算机视觉领域物体检测算法的重要性能评价指标,长期用于相关学术研究成果的定量比较.我们在实践中发现AP指标存在缺陷不能准确反映物体检测算法的数量准确性.我们具体分析了AP存在缺陷的两种表现,首先是AP定义和计算方法允许虚警出现时召回率不变,出现一个召回率对应多个准确度的情况,AP值计算仅和该召回率对应的最大准确度有关,使得准确度降低的虚警检测被AP忽略或者称为被AP隐藏.其次,我们发现在高召回率区域,AP对于准确度的增量和对于召回率的增量偏好不同,且偏向于召回率增加,这意味着检测算法增加虚警对增加AP值更有益.AP值评价指标的缺陷有很大的危害,首先导致该领域研究存在人为反复调整阈值从而获得高AP值的可能,其次能够增强物体检测数量准确性的研究方法被长期抑制,阻碍了该领域的创新发展,最后导致追求高AP值的学术研究与追求数量准确性的实际应用严重分离,很多学术成果在实际运用中无法达到应用的要求.基于上述分析,我们呼吁对AP指标进行批判式的改进,以弥合学术研究和实际需求的鸿沟,促进技术创新和产业运用良性互动.

Investigations of 3D Surface Roughness Characteristic＇s Accuracy

The existing surface roughness standards comprise only two dimensions. However, the real roughness of the surface is 3D （three-dimensional）. Roughness parameters of the 3D surface are also important in analyzing the mechanics of contact surfaces. Problems of mechanics of contact surfaces are related to accuracy of 3D surface roughness characteristic. One of the most important factors for 3D characteristics determination is the number of data points per x and y axes. With number of data points we understand its number in cut-off length. Number of data points have substantial influence on the accuracy of measurement results, measuring time and size of output data file （especially along the y-axis direction, where number of data points are number of parallel profiles）. Number of data points must be optimal. Small number of data points lead to incorrect results and increase distribution amplitude, but too large number of data points do not enlarge range of fundamental information, but substantially increase measuring time. Therefore, we must find optimal number of data points per each surface processing method.