Benchmarking stereo reconstruction for 3D printable Martian terrain models
Researchers evaluated a pipeline for creating 3D printable models from Mars rover images, facing challenges with low-texture and irregular Martian terrain. They found that while RAFT-Stereo improved accuracy on standard benchmarks, its performance degraded on actual Martian imagery due to weaker edge alignment. The study also highlighted trade-offs in geometry completion methods, with alpha shapes preserving detail but fragmenting structure, and Poisson reconstruction creating more coherent meshes at the cost of unsupported surfaces. AI
IMPACT Demonstrates limitations of current AI stereo reconstruction methods for specialized, low-texture environments, suggesting a need for domain-specific validation.