项目简介
LingBot-Map 是一种前馈式 3D 基础模型,能从流式数据重建场景。它采用几何上下文 Transformer 统一坐标定位、密集几何线索和漂移校正,通过分页 KV 缓存实现约 20 FPS 的长序列推理。
README 预览
\n \n\nLingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction\n\nRobbyant Team\n\n\n\n\n\n[](https://arxiv.org/abs/2604.14141)\n[](lingbot-map_paper.pdf)\n[](https://technology.robbyant.com/lingbot-map)\n[](https://huggingface.co/robbyant/lingbot-map)\n[](https://www.modelscope.cn/models/Robbyant/lingbot-map)\n[](LICENSE.txt)\n\n\n\nhttps://github.com/user-attachments/assets/fe39e095-af2c-4ec9-b68d-a8ba97e505ab\n\n-----\n\n### 🗺️ Meet LingBot-Map! We've built a feed-forward 3D foundation model for streaming 3D reconstruction! 🏗️🌍\n\nLingBot-Map has focused on:\n\n- **Geometric Context Transformer**: Architecturally unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single streaming framework through anchor context, pose-reference window, and trajectory memory.\n- **High-Efficiency Streaming Inference**: A feed-forward architecture with paged KV cache attention, enabling stable inference at ~20 FPS on 518×378 resolution over long sequences exceeding 10,000 frames.\n- **State-of-the-Art Reconstruction**: Superior performance on diverse benchmarks compared to both existing streaming and iterative optimization-based approaches.\n\n---\n\n## 📑 Table of Contents\n\n\nClick to expand\n\n- [📰 News](#-news)\n- [📋 TODO](#-todo)\n- [⚙️ Installation](#️-installation)\n- [📦 Model Download](#-model-download)\n- [🚀 Quick Start](#-quick-start)\n- [🎬 Interactive Demo (`demo.py`)](#-interactive-demo-demopy)\n - [Try the Example Scenes](#try-the-example-scenes)\n - [Streaming with Keyframe Interval](#streaming-with-keyframe-interval)\n - [Windowed Inference (for long sequences, >3000 frames)](#windowed-inference-for-long-sequences-3000-frames)\n - [Sky Masking](#sky-masking)\n - [Visualization Options](#visualization-options)\n - [Performance & Memory](#performance--memory)\n- [🎥 Offline Rendering Pipeline (`demo_render/batch_demo.py`)](#-offline-rendering-pipeline-demo_renderbatch_demopy)\n- [📜 License](#
常见问题 (2)
informational为什么论文中提到13个Oxford Spires场景,而列出的序列总数却是14个?实际评估的是哪些序列?
"13 scenes" 的描述是一处笔误:该列表包含 14 个序列。作者最初评估了全部 14 个序列,得到 ATE RMSE = 5.59,但发现其中 4 个场景的位姿与点云未对齐。因此,最终论文报告了 10 个场景的结果,ATE RMSE = 6.42。
排除的序列:
- blenheim-palace-01: 4.116
- blenheim-palace-02: 3.687
- blenheim-palace-05: 2.567
- christ-church-01: 3.6793
包含的序列及可用的逐序列 ATE 值(前 320 帧,每 12 帧采样一次):
- bodleian-library-02: 9.592
- christ-church-02: 3.2969
- christ-church-03: 0.4530
- christ-church-05: 35.987
- keble-college-02: 3.016
- keble-college-03: 1.957
- keble-college-04、keble-college-05、observatory-quarter-01 和 observatory-quarter-02 也包含在内,但它们各自的数值在讨论帖中被截断。
作者表示,场景描述将在论文的下一版本中更正。
故障排除哪个检查点用于复现 Oxford Spires 表 2 的结果?
对于 Table 2 的 Oxford Spires 结果,请使用 **lingbot-map.pt 的第 160 个 epoch 检查点**。维护者表示,lingbot-map.pt 和 lingbot-map-long.pt 都能取得比 Table 2 所报告数值更好的 ATE,但 lingbot-map-long.pt 并非该表格所使用的检查点。对于 AUC,请遵循 **DA3 pose-AUC 计算方法**。该讨论串中并未提供具体的图像预处理、裁剪/缩放、内参处理方式以及官方评估代码;维护者表示评估代码将会发布,因此请使用该官方代码来解决剩余的 AUC 差距。