#1212th of 23 in Open models
lingbot-map
Feed-forward streaming 3D reconstruction and camera pose model for video
- Stars
- 17.5k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
LingBot-Map is a PyTorch model, the Geometric Context Transformer, that reconstructs 3D point clouds and camera poses from a video or image folder in a streaming fashion. It uses paged KV cache attention via FlashInfer, with an SDPA fallback, and the README reports about 20 FPS at 518×378 over sequences beyond 10,000 frames. It ships an interactive viser viewer, a windowed mode for long inputs, and an offline renderer that produces flythrough MP4s.
Who it is for: Researchers and engineers doing 3D reconstruction or pose estimation from video
Strengths
- Streaming inference reported at ~20 FPS on 518×378 over sequences above 10,000 frames
- Windowed mode and keyframe interval handle very long videos, e.g. a ~25,000-frame demo
- Includes evaluation scripts for KITTI and Oxford Spires, plus an interactive viser viewer
- Checkpoints on Hugging Face and ModelScope under Apache-2.0
Weaknesses
- Quality degrades when the KV cache holds more than 320 views, so long videos need tuning
- Offline renderer needs Kaolin, ffmpeg and a manually built CUDA extension
- Pinned to PyTorch 2.8.0 with CUDA 12.8 for prebuilt Kaolin wheels
- No GPU memory figures stated; low-VRAM use points to a third-party fork
What it needs
- GPU required
- Needs PyTorch 2.8.0 (CUDA 12.8), FlashInfer (recommended, SDPA fallback), viser, onnxruntime, NVIDIA Kaolin (offline renderer), ffmpeg (offline renderer), open3d (offline renderer)
- Models: lingbot-map checkpoint (.pt), lingbot-map-stage1 checkpoint
- port 8080
Also in Open models
See all 23| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 86 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 83 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 74 out of 100 |
| #4 | unlimited-ocrOCR model that parses long documents and multi-page PDFs in one pass | 69 out of 100 |
| #5 | speechPyTorch toolkit for training and deploying ASR, TTS and speech LLM models | 67 out of 100 |