Instance Segmentation of Gait Silhouettes from a Remote Sensing Perspective

This article studies the methods and models of gait silhouette instance segmentation from a remote sensing perspective in recent years, and addresses their shortcomings by making some improvements, thereby enhancing accuracy and practicality. Despite annual upgrades to instance segmentation models, their accuracy in segmenting gait silhouettes on remote sensing equipment remains unsatisfactory, especially for outdoor datasets. This paper investigates the methods and models of gait silhouette instance segmentation and makes certain improvements based on SOTA (state-of-the-art) models. The model innovatively uses DenseNet as the backbone for the FPN layers within the Mask R-CNN framework and replaces the traditional BN layers with FRN layers and TLU. The results show that our model achieves significantly better performance than previous instance segmentation models under remote sensing conditions.