lingbot-map: Real-time 3D reconstruction of videos with over 10,000 frames at 20 FPS

It reconstructs 3D scenes in real time while correcting positional errors and drift that occur during long video streaming. It improves precision by integrating coordinate reference points, pose reference windows, and trajectory memory into a single streaming framework.

With a Feed-forward structure applying Paged KV Cache Attention, it stably processes long sequences of over 10,000 frames at approximately 20 FPS at a resolution of 518x378. It produces results in a single inference pass without iterative optimization.

It demonstrated superior performance compared to existing streaming and iterative optimization-based methods on various benchmarks such as KITTI and Oxford Spires. It shows robust results even when processing long-duration data, such as 13-minute indoor walkthrough videos.

It enables real-time visualization via a browser-based viewer and also provides automatic sky region removal. It is suitable for tasks where real-time performance is critical, such as large-scale 3D mapping or autonomous driving data preprocessing.