Spaces:

---

NicolasCC

/

LongStream

like 1

Running on Zero

LongStream is an academic research project accompanying our CVPR 2026 paper on streaming 3D reconstruction and pose estimation from long monocular videos. The key idea is to reconstruct scenes and estimate camera motion in a strictly-online setting, where frames are processed sequentially without access to future observations. This setting is important for real-world applications such as robotics, AR/VR, and embodied perception systems.

We are applying for GPU support primarily to build a public demonstration of the method. The demo will allow users to upload or select video sequences and visualize the incremental reconstruction process, including streaming pose estimation and evolving 3D geometry. This will help researchers and practitioners better understand the challenges of real-time, future-invisible reconstruction.

The GPU resources will be used to run inference for interactive demos and example sequences, enabling the community to explore the behavior of streaming reconstruction systems on long videos.

The project will be released with open code, demo examples, and documentation, aiming to contribute an accessible showcase of real-time 3D scene understanding for the computer vision community.