Revisiting the Domain Gap Issue in Non-cooperative Spacecraft Pose Tracking

The deep learning (DL) algorithms have emerged as the foremost approach for close-range navigation of non-cooperative spacecraft. Given the unavailability of in-orbit images, DL models are typically trained on synthetic data. However, when deployed in real-world scenarios, they often encounter a domain gap that leads to performance degradation. To address this, we propose a self-supervised framework based on RANSAC EPnP. Specifically, we first trained a landmark regression network and an object detection network on synthetic data. Utilizing the trained landmark regression network, we then infer keypoints on real-world images. Through RANSAC EPnP, we filter outliers and calculate poses as pseudo-labels. Building on this, the pose estimation network is further trained, optimizing outliers to bridge the domain gap. The proposed method brings a significantly lower training cost compared to adversarial training, the prevailing method for bridging the domain gap, making it suitable for in-orbit training. Moreover, we utilize a Kalman filter to predict the bounding boxes, which circumvents the domain gap’s impact on the performance of the object detection network, resulting in more precise bounding boxes. Lastly, we validated the performance of the proposed algorithm on the SPEED+ and SPARK 2024 datasets, achieving the 2nd place in the SPARK 2024 competition.