MVDiffusion++: A Revolutionary Approach to 3D Object Reconstruction Without Camera Poses
核心概念
Pose-free architecture enables high-resolution 3D object reconstruction without camera poses.
摘要
Presents MVDiffusion++, a neural architecture for 3D object reconstruction.
Utilizes self-attention among 2D latent features for 3D consistency without camera poses.
Introduces a view dropout strategy to reduce training-time memory footprint.
Outperforms current state-of-the-art methods in novel view synthesis and 3D reconstruction metrics.
Combines with text-to-image generative model for text-to-3D application.
MVDiffusion++
統計資料
MVDiffusion++ achieves state-of-the-art performance on single-view reconstruction with IoU of 0.6973 and Chamfer distance of 0.0165 on Google Scanned Objects dataset.
引述
"Our surprising discovery is that self-attention among 2D latent image features is all we need for 3D learning without projection models or camera parameters."
"MVDiffusion++ significantly outperforms the current state of the arts in novel view synthesis and sparse-view reconstruction."