Bridging the domain gap in spacecraft pose estimation: A novel image generative model

With the increasing complexity of space exploration missions, vision-based autonomous spacecraft navigation and servicing technologies are increasingly important. However, synthetic spacecraft imagery used for training deep learning models often introduces a significant domain gap with real-world data, limiting model generalization. We propose a sequence-conditioned image translation framework that combines a conditional Generative Adversarial Network (GAN) with a Convolutional Long Short-Term Memory (ConvLSTM) module to reduce this domain gap. This approach generates high-fidelity spacecraft images, bridging the domain gap. The model integrates a residual-enhanced U-Net generator with a PatchGAN discriminator, enhancing image realism, while the ConvLSTM module captures sequential features to ensure consistency across frames. Experimental results show that the framework markedly narrows the synthetic-to-real domain gap: the Fréchet Inception Distance (FID) drops from 419.4 to 40.1, the Structural Similarity Index (SSIM) rises from 0.286 to 0.917, and the Peak Signal-to-Noise Ratio (PSNR) increases from 27.8 dB to 34.6 dB. The concurrent boost in pose-estimation accuracy further corroborates the method’s effectiveness, enabling synthetic imagery to deliver near-real-world performance in spacecraft pose estimation and autonomous mission scenarios while reducing reliance on real datasets.