A Dynamic Segmentation Model for Spacecraft with Data Generation

With the development of space technology, intelligent techniques attract attentions in the aerospace spheres. However, the scarcity of space datasets presents challenges for the development of perception models based on deep learning. In this paper, we first incorporate orbital dynamic constraints to design a spacecraft semantic data generation workflow and provide a data interface for the segmentation tasks. Furthermore, we devise a dynamic training strategy for spacecraft segmentation models based on "data generating", achieving relative synchronization between data generation and model training. Finally, we demonstrate the efficacy of the proposed model through extensive comparative experiments. The final model achieves 83.48 mIoU and 89.79 mAcc with only 49.34 MB parameters, 40.82 GFLOPS and 12.98 FPS on the platform.