In-Orbit MapAnything: An Enhanced Feed-Forward Metric Framework for 3D Reconstruction of Non-Cooperative Space...

Area of Science:

• Robotics and Space Engineering
• Computer Vision
• Photogrammetry

Background:

• Accurate 3D reconstruction of space targets is crucial for active debris removal and on-orbit servicing.
• Extreme space lighting, sparse textures, and specular reflections challenge current reconstruction methods, degrading accuracy.
• General feed-forward models struggle with domain shifts encountered in space environments.

Purpose of the Study:

• To develop an enhanced 3D reconstruction framework, In-Orbit MapAnything, specifically designed for the challenging space environment.
• To address data scarcity and improve feature robustness against specular reflections in space target reconstruction.

Main Methods:

• Construction of a high-quality space target dataset with extreme illumination and auxiliary modalities.
• Introduction of the SatMap-Adapter module with hierarchical cascade sampling and adaptive fusion for feature enhancement.
• Application of weight-decomposed low-rank adaptation for parameter-efficient fine-tuning of the pre-trained backbone.

Main Results:

• Significant reduction in absolute relative error (15.23%) and Chamfer distance (20.02%) compared to the baseline MapAnything model.
• Effective suppression of reconstruction noise on metallic surfaces and recovery of fine geometric structures.
• Maintenance of rapid inference speed, demonstrating practical applicability.

Conclusions:

• In-Orbit MapAnything provides a robust and accurate solution for 3D reconstruction in extreme space conditions.
• The proposed feature-enhanced framework effectively addresses limitations of existing methods for space applications.
• This advancement supports critical space operations like debris removal and on-orbit servicing.