Domain-adaptive Object Detection with Dynamic Mosaic and Consistent Learning
Ding Gao, Jian Yang · 2023
In recent years, the teacher-student model has made breakthrough progress in the domain-adaptive object detection problems. However, the consistent learning and data augmentation mechanisms in this framework are not perfect. The traditional strong data augmentation mechanisms would usually make the entire target objects to be occluded. Therefore, the original targets may disappear in the augmented image. However, the pseudo-labels contain these occluded object boxes, which would hinder the model learning. This paper proposes a novel dynamic mosaic module. This module generates variable-sized mask blocks by using the information entropy of the pseudo-labels, which will ensure the targets in complex scenes not to be over-occluded during consistency learning, and achieve effective data augmentation in simple scenes. This method can dynamically adapt to scenes of different levels of complexity, enhancing the model’s adaptability in different fields.