Pseudo-label generation method, and source-free domain scenario adaptive occlusion-aware seamless segmentation method and system

Abstract

Disclosed in the present invention are a pseudo-label generation method, and a source-free domain scenario adaptive occlusion-aware seamless segmentation method and system. In the pseudo-label generation method, threshold filtering and data volume comparison are performed on conventionally used instance-level pseudo labels, thereby further improving the marking precision of the pseudo labels; moreover, the generated pseudo labels are used, and an uncertain-region-guided weighted loss for an instance-level prediction branch in an occlusion-aware seamless segmentation task is designed, thereby improving the accuracy of a segmentation model; and in combination with filtering of low-quality pseudo labels, a modality-agnostic-guided instance mixing strategy is proposed, thereby further increasing the number of samples available for training. Therefore, the problem of a poor segmentation effect of a finally trained model caused by the scarcity of samples in certain categories is solved, and the source-free domain scenario adaptive effect is finally improved.

Images