Construction Methods of Defective datasets for Deep Learning Vision Algorithms

In response to the lack of defective datasets for deep learning vision algorithms, we researched methods for constructing such datasets. We analyzed the reasons behind the failures of deep learning vision algorithms and summarized 18 types of defects that lead to these failures. These defects were categorized into four major types: dataset, algorithm training, algorithm inference, and algorithm deployment. We constructed 266 defective datasets based on six typical vision tasks: image classification, object detection, object tracking, object extraction, semantic segmentation, and image fusion. Through comparative experiments, we verified the impact of these defective datasets on the performance of various vision algorithms.