Honey, I Shrunk the Arc de Triomphe!
Researchers have developed a new dataset called MetricScenes to address the "scale-collapse" issue in monocular geometry estimation, where distant objects are inaccurately represented. This dataset, compiled from internet photos and stereo imagery, provides metrically-grounded, in-the-wild scenes. Fine-tuning the MoGe-2 model on MetricScenes significantly improves its accuracy in estimating absolute scale for unconstrained environments. AI
IMPACT Improves 3D scene understanding from monocular images, potentially aiding applications in robotics and augmented reality.