A New Dataset and Performance Benchmark for Real-time Spacecraft Segmentation in Onboard Flight Computers

Second, no established benchmarks evaluate performance under hardware constraints equivalent to flight computers (less than 4GB RAM, CPU-only inference) and inference time constraints (inference time less than 0. 95 second). Third, conventional metrics such as the Dice coefficient fail to capture boundary localization precision critical for proximity operations. These gaps hinder the development of algorithms that can be deployed on resource-constrained orbital platforms. To address these challenges, we introduce a dual-methodology dataset synthesis approach. Building on the Pose-Bowl and Spacecrafts datasets, we created our dataset, Spacecraft With Masks (SwiM), through two complementary strategies: (1) superimposing existing spacecraft images on augmented open-source backgrounds with photometric/geometric distortions to mimic real-world noise and distortions in image acquisition in space, and (2) generating synthetic samples via NASA's TT ALOS (Toolset for Training and Labeling in an Optical Simulator) pipeline, which integrates astrophysical backgrounds generated using stable diffusion with procedu-rally rendered 3D spacecraft models. This hybrid methodology achieves unprecedented diversity in spacecraft poses, lighting conditions, and environmental contexts while maintaining physical precision and simulating camera distortions and noise. To the best of our knowledge, our SWiM dataset, consisting of nearly 64k images with annotations, is the largest and most comprehensive spacecraft segmentation dataset to date.