SUFLECA Enhances CAD-to-Image Alignment
Saad Ejaz · hf · 2026-07-17
SUFLECA proposes a weakly supervised zero-shot framework for CAD-to-image alignment, aiming to estimate an object's 9D pose (rotation, translation, and non-uniform scaling) from a single RGB image. The authors note that existing zero-shot methods mostly rely on vision foundation models for appearance matching, which are prone to failure under occlusion and sim-to-real domain shifts.
Methodologically, SUFLECA features two key designs:
- Geometry-aware feature learning: Utilizes 674,000 images across 12 real and synthetic datasets, learning compact geometric features from pre-trained visual representations supervised by NOCS.
- Geometry-consistent matching: Proposes a matching algorithm that establishes more reliable one-to-one CAD-image correspondences.
Experimental results show that this method can achieve sub-second alignment on a single instance without requiring iterative pose optimization. On ScanNet25k, it achieved 33.4% / 42.3% class/instance accuracy, outperforming the strongest zero-shot baseline by 10.3 / 12.2 percentage points, and is the first to surpass fully supervised methods on this benchmark. Code is open source.
More from Embodied
- New survey maps how agentic systems are learning to improve themselves — SchmidhuberAI · 2026-07-21
- ReViV reconstructs egocentric 4D viewer-and-scene dynamics from one monocular video — ethz-vlg · 2026-07-21
- Neuracore pitches a one-loop platform for robotics data, training, and deployment — stepjamUK · 2026-07-21
- Humanoid robots are moving from labs into public culture — Olivier__OG · 2026-07-21
- WAIC panel says world models still lack consensus on routes, metrics, and deployment — 量子位 · 2026-07-21
- Polymarket puts Tesla’s California robotaxi launch odds at 16% this year — Polymarket · 2026-07-21