Can we trust bounding box annotations for object detection

Most methods of target detection need a large number of annotated data sets for training and evaluation . The most common representation is the bounding box （BBs）, It is usually defined as the smallest area rectangle that contains the entire object area . However , The annotation process is somewhat subjective （ Especially when there is occlusion ）, When the annotator is tired , Its quality may decline .

In this paper , The theoretical and experimental results provided by us show that , In popular data sets for object detection , Even if there are small annotation differences ,IoU Will also be greatly affected . therefore , It is usually used to evaluate the average accuracy of the object detector （AP） Values can also be affected by annotation bias or noise , Especially for small objects and more stringent IoU threshold .

Related work

Some papers have evaluated bias and annotation errors in large image classification datasets （ or “ problem ”）.Tsipras wait forsomeone [23] Large image classification datasets are evaluated （ Such as ImageNet[1]） Comment deviation in , And how the data set creation process may lead to deviation . They point out that there are multiple valid tags and ambiguous categories of images , For example, missiles / The shells . about ImageNet, stay [18,19] Similar findings on category fuzziness are reported in , This may affect the top-1 Accurate quality indicators .