LingBot-Map Tutorial: GPU-Aware Inference and Point Cloud Export
LingBot-Map
This tutorial demonstrates an end-to-end streaming 3D reconstruction pipeline using LingBot-Map. It covers GPU detection, automatic tuning of inference parameters, model setup, mixed-precision inference, point cloud export in PLY, NPZ, or GLB formats.
The tutorial implements a complete streaming 3D reconstruction pipeline with LingBot-Map. It begins by configuring input source, reconstruction settings, checkpoint selection, and output controls. The pipeline probes the GPU and automatically adjusts frame limits, camera iterations, scale frames, and KV-cache parameters based on detected VRAM. It then installs repository dependencies, downloads the pretrained checkpoint, preprocesses frames, and builds the GCTStream model with streaming attention and long-range trajectory memory. Mixed-precision inference decodes camera poses and intrinsics, converts depth maps to world-coordinate point clouds, validates geometry, and visualizes the scene. Results are exported as PLY, NPZ, or GLB files.
- Abbreviations
- GPU = Graphics Processing Unit — графический процессор
- VRAM = Video Random Access Memory — видеопамять
- KV = Key-Value — ключ-значение
- PLY = Polygon File Format — формат файла полигонов
- NPZ = NumPy Zipped Archive — сжатый архив NumPy
- GLB = GL Binary — бинарный формат GL
- CUDA = Compute Unified Device Architecture — архитектура параллельных вычислений
- FPS = Frames Per Second — кадры в секунду
Source: MarkTechPost —
original