Six chapters, one thread六个章节，一条主线

I

PhD research main line博士生涯研究主线

Building & Spatial Structure Understanding with Offset Tokens基于偏移量词元的建筑与空间结构理解

My PhD research targets building extraction under off-nadir satellite photography. I first proposed the concept of the Offset Token, which brings off-nadir photography into the perceptual scope of large vision models. Around this concept I built a complete loop — vectorised building extraction, positional repair between historical vector maps and updated imagery, efficient optimisation of the offset token, and full decoupling of offset learning.博士期间的研究方向为遥感卫星偏移摄影视角下的建筑物提取。研究主体上， 首先提出了偏移量词元（Offset Token）的概念，将偏移摄影带入大模型的理解视角； 并围绕所提出的偏移量词元概念展开更加细腻的研究，实现了偏移摄影下矢量化的建筑物提取、 历史矢量地图与更新影像之间的位置修复、偏移量词元的高效优化计算、偏移量学习解耦等内容， 完成了「概念提出 → 具象化应用 → 任务解耦与独立 → 底层理论降维」的闭环。

@article{li2024obm,
  title   = {Prompt-Driven Building Footprint Extraction in Aerial Images
    with Offset-Building Model},
  author  = {Li, Kai and others},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year    = {2024},
  doi     = {10.1109/TGRS.2024.3487652}
}

@article{li2025polyfootnet,
  title   = {PolyFootNet: Extracting Polygonal Building Footprints in
             Off-Nadir Remote Sensing Images},
  author  = {Li, Kai and others},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year    = {2025}
}

@article{li2025dragosm,
  title  = {DragOSM: Extract Building Roofs and Footprints from Aerial Images
            by Aligning Historical Labels},
  author = {Li, Kai and Weng, X. and Deng, Yupeng and Meng, Yu and Pang, C.
            and Xia, Gui-Song and Zhao, Xiangyu},
  journal = {arXiv preprint arXiv:2509.17951},
  year   = {2025}
}

@article{li2026oblicity,
  title  = {ObliCity: A Benchmark and Baseline for Roof-to-Ground Projection
            Displacement Correction},
  author = {Li, Kai and Deng, Yupeng and Deng, L. and Xi, Z. and Wang, C.
            and Zhang, J. and Ji, Y. and Meng, Yu and Zhao, Xiangyu},
  journal = {arXiv preprint arXiv:2607.25210},
  year   = {2026}
}

@article{li2025lodeot,
  title  = {LODEOT: Low-Dimensional and Efficient Offset Tokens for Building
            Footprint Extraction from Off-Nadir Imagery},
  author = {Li, Kai and others},
  year   = {2025}
}

II

Spatial understanding空间维度

Multimodal LLMs for 3D Scene Understanding, with Reinforcement Learning多模态大模型驱动的 3D 场景理解与强化学习

After accumulating substantial experience in 3D spatial understanding, this line extends my work towards today's frontier: letting large multimodal models reason about 3D structure directly, without camera calibration or dense views — and, building on that foundation, using reinforcement learning to keep sparse-view indoor reconstruction consistent across viewpoints.在积累了丰富的三维空间理解经验之后，这一部分把工作拓展到当下最前沿的方向： 让多模态大模型在没有相机参数、没有密集视图的条件下直接推理三维结构；并在此基础上， 尝试用强化学习让稀疏视角下的室内场景重建保持跨视角一致。

@article{li2026scenix,
  title  = {Scenix: Sparse-View 3D Scene Reconstruction via Executable
            Scene Programs},
  author = {Li, Kai and Jiang, L. and Li, Zhenyang and Dong, J. and Zhang, J.
            and Yin, Y. and Zhang, Runze and Yan, K. and Huang, X. and others},
  journal = {arXiv preprint arXiv:2608.07012},
  year   = {2026}
}

III

Ongoing research在研方向

Current Focus: Memory in Video World Models当前重心：视频世界模型中的记忆机制

Video world models learn to predict how a scene evolves, but they largely lack a persistent notion of what they have already seen. My current work asks how such a model should remember: how to retain scene state over long horizons, stay consistent when the camera revisits a place, and keep memory affordable as the rollout grows. This is the temporal counterpart to the spatial understanding I built during my PhD.视频世界模型能够预测场景如何演化，却普遍缺少对「已经看到过什么」的持久表征。 我目前的研究关注这类模型应当如何记忆：如何在长时序上保持场景状态、 如何在镜头重新回到同一位置时维持一致、以及如何在推演变长时控制记忆开销。 这与我博士期间构建的空间理解能力，构成时间维度上的呼应。

Other Collaborative Publications其他合作论文