TALON: Token-Aligned Lightweight Adapters for 6-DoF Spacecraft Pose Estimation
Researchers have developed TALON, a novel method for estimating the 6-DoF pose of spacecraft using monocular vision. TALON injects spatiotemporal 3D adapters into a frozen Vision Transformer (ViT) and employs a patch-token alignment loss to ground features to keypoint structures. This approach significantly enhances performance by incorporating temporal information without requiring full backbone fine-tuning or auxiliary optical flow networks, adding less than 5% parameters to the original model. AI
IMPACT Enhances spacecraft pose estimation accuracy and efficiency by leveraging temporal data and lightweight adapters within Vision Transformers.