PASTEL: Panoramic Alignment for Monocular 4D Scene Reconstruction

Paper

Yuankun Yang¹, Yi Wei², Bo Bai², Wenyang Zhou², Li Zhang¹ ✉

¹School of Data Science, Fudan University    ²Central Media Technology Institute, Huawei

ECCV 2026

PASTEL targets monocular 4D scene synthesis beyond visible camera limits. It aligns the scene into a panoramic representation, plans adaptive extrapolation trajectories, and strategically exploits generative priors to synthesize invisible regions while preserving consistency with the input video.

Method

PASTEL reformulates 4D scene synthesis through panoramic alignment:

• Panoramic scene alignment maps back-projected observations into a unified panoramic domain with explicit visibility boundaries.
• Adaptive trajectory identification selects expansion directions that maximize out-of-view exploration with minimal deviation from the original cameras.
• Comprehensive view expansion & strategic supervision warps static and dynamic content along designed trajectories, then distills generative outputs only into invisible and unreliable regions.

Given a monocular video, PASTEL first performs panoramic representation alignment and trajectory identification, then conducts static–dynamic view expansion and strategic 4D synthesis guided by generative priors.

BibTeX

If you find this project helpful, please consider citing our paper:

@inproceedings{yang2026pastel,
  title     = {{PASTEL}: Panoramic Alignment for Monocular 4D Scene Reconstruction},
  author    = {Yang, Yuankun and Wei, Yi and Bai, Bo and Zhou, Wenyang and Zhang, Li},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}