UGG：统一生成式几何模型整合视频生成与深度估计

摘要

UGG 将视频生成、深度估计和相机位姿预测合并为单一扩散模型，把缺失的视觉与几何信息当作待生成内容，从而从视频重建场景几何、按相机运动生成视频，并在多种 3D 解释中搜索最一致的一种。在相机位姿估计上，UGG 在 ScanNet++ 上的轨迹误差低于 Geo4D；在多个基准上也优于专用视频生成模型，尤其在观测有限时。

摘要由模型根据原文撰写，以原文为准。

"UGG: Unified Generative Geometry Model"

Video generation models don't necessarily understand the 3D geometry behind them, since video gen, depth estimation, and camera pose prediction are mostly treated as separate tasks.

This paper introduces UGG, which combines all three tasks into a single diffusion model. So by treating missing visual and geometric information as something to generate, UGG can reconstruct scene geometry from videos, generate videos from camera movements, and even search through multiple possible 3D interpretations to find the most consistent one.

On camera pose estimation, UGG reduces trajectory error better than Geo4D on ScanNet++. It also outperforms specialized video generation models on several benchmarks, especially when only limited observations are available.

https://www.alphaxiv.org/abs/2611.14440

UGG：统一生成式几何模型整合视频生成与深度估计

摘要

正文 · 原文

引用原文