arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2603.01010cs.CV
收藏

GeodesicNVS: 基于概率密度几何流的新型视角合成

GeodesicNVS: Probability Density Geodesic Flow Matching for Novel View Synthesis

Huawei Hilbert Research Center (Dresden)(华为希尔伯特研究中心(德累斯顿));Technical University of Munich(慕尼黑工业大学);Karlsruhe University of Applied Sciences(卡尔斯鲁厄应用科学大学)

Xuqin Wang, Tao Wu, Yanfeng Zhang, Lu Liu, Mingwei Sun, Yongliang Wang, Niclas Zeller, Daniel Cremers

AI总结:

本文提出PDG-FM方法,通过数据依赖的几何正则化提升确定性流匹配的视角一致性,实验证明在Objaverse和GSO30数据集上优于扩散模型。

摘要:

近期生成模型的进步显著提升了新型视角合成,但保持多视角一致性仍具挑战。扩散模型依赖随机噪声到数据的转换,会掩盖确定性结构并导致不一致的视角预测。本文提出数据到数据的流匹配框架,学习配对视角间的确定性转换,通过显式数据耦合增强视角一致性合成。基于此,我们提出概率密度几何流匹配(PDG-FM),将插值轨迹对齐数据流形上的密度基几何流。为实现可 tractable 的几何估计,我们采用教师-学生框架,将基于密度的几何插值器蒸馏为高效的 ambient-space 预测器。实验表明,我们的方法在 Objaverse 和 GSO30 数据集上优于扩散基线,展示了在确定性流匹配中引入数据依赖几何正则化以实现一致新型视角生成的优势。

英文摘要:

Recent advances in generative modeling have substantially enhanced novel view synthesis, yet maintaining consistency across viewpoints remains challenging. Diffusion-based models rely on stochastic noise-to-data transitions, which obscure deterministic structures and yield inconsistent view predictions. We advocate a Data-to-Data Flow Matching framework that learns deterministic transformations between paired views, enhancing view-consistent synthesis through explicit data coupling. Building on this, we propose Probability Density Geodesic Flow Matching (PDG-FM), which aligns interpolation trajectories with density-based geodesics of a data manifold. To enable tractable geodesic estimation, we employ a teacher-student framework that distills density-based geodesic interpolants into an efficient ambient-space predictor. Empirically, our method surpasses diffusion-based baselines on Objaverse and GSO30 datasets, demonstrating improved structural coherence and smoother transitions across views. These results highlight the advantages of incorporating data-dependent geometric regularization into deterministic flow matching for consistent novel view generation.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑