A comprehensive, end‑to‑end benchmark suite for evaluating 3D Geometric Foundation Models (GFMs) on effectiveness, robustness, and efficiency across diverse tasks and data domains.
-
Rigorous Effectiveness and Efficiency Evaluation
-
Five Core Tasks
- Sparse‑view Depth Estimation
- Monocular Video Depth Estimation
- Multi‑view 3D Reconstruction
- Multi‑view Relative Pose Estimation
- Novel View Synthesis
-
Diverse Dataset Support
- Standard indoor/outdoor and object‑centric benchmarks
- Challenging out‑of‑distribution scenarios (drone footage, dynamic scenes, air‑ground pairs)
-
Reproducible Evaluation Toolkit
- Standardized data loaders and preprocessing
- Unified metric implementations (Acc, Comp, NC, PSNR, SSIM, latency, memory)
- Automated alignment and scale‑normalization pipelines
-
Modular & Extensible
- Plug‑and‑play support for any feed‑forward or diffusion‑based GFM
- Clear APIs for adding new tasks, metrics, or datasets