Reproduce “Emergent Outlier View Rejection in Visual Geometry Grounded Transformers” on OpenResearch

About this paper

Jisang Han, Sunghwan Hong, Jaewoo Jung, Wooseok Jang, Honggyu An, Qianqian Wang, Seungryong Kim, Chen Feng

RobustVGGT filters distractor images before repeating VGGT reconstruction, using final-layer attention or feature similarity scores from an initial pass. It retains views above fixed thresholds without retraining or changing the architecture. On Phototourism, RobustVGGT-F reports mean ATE of 0.2650 versus 0.3504 for VGGT across three noise settings; exceptions on other benchmarks preclude a universal advantage.

Key insights

• State-of-the-art feed-forward 3D reconstruction transformers (e.g., VGGT) inherently develop an emergent capability for geometry-aware outlier suppression, without explicit training or modules for this task.
• This emergent behavior is most pronounced in the final layers of the transformer, where the model's geometric reasoning refines its understanding of view consistency.
• Leveraging these internal, emergent signals from existing models can provide a highly effective, training-free, and efficient solution for robustifying 3D reconstruction against noisy inputs.

Summary by alphaXiv.