画像から 3D 生成と三次元再構成
一枚または複数の写真から三次元構造を復元する
本ページの本文は英語で提供されています。タイトルと導入は日本語化されています。
この能力とは何か
Takes one or several images and outputs a 3D representation — mesh, point cloud, depth map or renderable radiance field. With multiple views the geometry is constrained jointly by the parallax across photos and is more trustworthy; with a single image, unobserved parts can only be inferred from priors. Unlike text-to-3D both geometry and appearance are anchored by the input images.
技術的にどう実現するか
The multi-view route estimates camera poses, triangulates, and refines dense depth and surfaces; neural radiance fields encode the scene as a differentiable volumetric field aligned to input views by differentiable rendering, and 3D Gaussian splatting represents it as oriented translucent ellipsoids that render faster. Single-image routes rely on priors learned from large 3D datasets, or on score distillation from a 2D diffusion prior.
代表的な製品
3関連する組織
代表的な用途
- Digital archiving of artefacts and buildings
- Spatial perception for robots and driving
- Turning product photos into 3D displays
- Digital twins of real locations for film and games
どう評価するか
- Chamfer distance
- Mean point distance between reconstructed and true surfaces
- F-score
- Balance of precision and recall within a distance threshold
- Novel-view PSNR / SSIM
- How close renderings from unseen views come to real photos
限界と難しさ
- Single-image reconstruction guesses entirely in unseen regions, and back surfaces appear invented once you rotate
- Reflective, transparent and textureless surfaces are hard to match, and reconstruction fails broadly there
- Output scale and metrics are inaccurate, so it cannot feed manufacturing or engineering measurement directly