Martine benchmark debuts for multi-view 3D surface reconstruction across viewpoints and lighting
ssh4net · x · 2026-11-11
Researchers led by Robin Bruneau and Lilian Calvet released Martine, a benchmark for multi-view 3D surface reconstruction that jointly evaluates methods across viewpoint coverage, resolution, and lighting in controlled acquisition setups.
Key points:
- Existing benchmarks restrict viewpoint, resolution, or illumination, biasing comparisons; Martine addresses this
- Dataset spans high geometric/optical diversity: low-texture regions, fine structures, transparent and specular materials, with curated challenge-level 3D masks
- Robotic arm + turntable gives dense viewpoint sampling; single-light and 30-LED multi-light regimes
- Paper highlights image–GT misalignment issues in OpenSubstance (COLMAP failures on an opaque flower and transparent glass)
- Dataset released under CC BY 4.0 with project page and paper
More from Multimodal
- Fish Audio launches Live 3, a full-duplex voice model that knows when to shut up — HeyAmit_ · 2026-11-11
- Seedream 5 nails darkfield microscopy: AI diatoms with glass shells and chloroplasts — CurieuxExplorer · 2026-11-11
- Early MiniMax H3 Image test: CAD-style landmark redesigns — CurieuxExplorer · 2026-11-11
- Using Codex with Capwise CLI to Auto-Generate a Product Intro Video — vista8 · 2026-11-11
- Creator blends H3Image with MiniMax H3 in experimental short music video — Mr_AllenT · 2026-11-11
- MAEP: training-free modality-aware pruning cuts MoE-MLLM expert memory over 90% — nota-ai · 2026-11-11