It is a common paradigm in object detection frameworks that the samples in training and testing have consistent distributions for the two main tasks: Classification and bounding box regression. This paradigm is popular in sampling strategy for training an object detector due to its intuition and practicability. For the task of localization quality estimation, there exist two ways of sampling: The same sampling with the main tasks and the uniform sampling by manually augmenting the ground-truth. The first method of sampling is simple but inconsistent for the task of quality estimation. The second method of uniform sampling contains all IoU level distributions but is more complex and difficult for training. In this paper, we propose an H+L-Sampling strategy, selecting the high and low IoU samples simultaneously, to effectively and simply train the branch of quality estimation. This strategy inherits the effectiveness of consistent sampling and reduces the training difficulty of uniform sampling. Finally, we introduce accurate detection confidence, which combines the classification probability and the localization accuracy, as the ranking keyword of NMS. Extensive experiments show the effectiveness of our method in solving the misalignment between classification confidence and localization accuracy and improving the detection performance.
展开▼
