[论文] VideoMaMa++: Temporally Consistent Video Matting via Preserve-and-Refi...

论文概要

研究领域: CV
作者: Sangbeom Lim, Seoung Wug Oh, Heeji Yoon, Seungryong Kim, Joon-Young Lee
发布时间: 2026-10-19
arXiv: 2610.26857

中文摘要

基于扩散的视频抠图利用视频扩散先验从有限合成数据实现强大的零样本泛化。然而庞大计算量将推理限制在6-24帧的短片段，普通视频会跨越多个片段。由于每个片段独立抠图，现有方法无法在片段边界保持一致性，产生闪烁边缘。我们提出掩码引导的扩散抠图框架：RGB和掩码条件作为独立token流通过完整时空自注意力交互；新颖的可学习保留token和精化token实现滑动窗口推理，已生成的抠图结果跨片段传播保持局部时间一致性，辅以全局共享参考锚点防止长序列误差累积。在标准基准和真实长视频上，我们的框架在细节准确性和时间一致性上均优于先前扩散方法。

原文摘要

Diffusion-based video matting has recently emerged as a leading approach, leveraging video diffusion priors from internet-scale pretraining to achieve strong zero-shot generalization from limited synthetic data. However, the heavy compute of these diffusion backbones limits inference to a short chunk of just 6 to 24 frames, so even ordinary videos span multiple chunks. Because each chunk is matted independently from coarse mask guidance, existing methods fail to maintain consistency across chunk boundaries, producing flickering edges and discontinuous transparency. We address this with a mask-guided diffusion-matting framework built on two ideas. The RGB and mask conditioning are represented as separate token streams that interact through full spatio-temporal self-attention, letting the mo...

自动采集于 2026-10-21

#论文 #arXiv #CV #小凯