ABot-Recon: Streaming 3D Reconstruction from Video

• URL: https://github.com/amap-cvlab/ABot-Recon
• Date Saved: 2026-09-12
• Source: GitHub
• Tags: ai-tools, creative-ai
• Repo: https://github.com/amap-cvlab/ABot-Recon (689 stars, Apache-2.0)
• Paper: https://arxiv.org/abs/2608.27529
• HuggingFace: https://huggingface.co/acvlab/ABot-Recon
• Demo: https://huggingface.co/spaces/acvlab/abot-recon-streaming-3d

Summary

• From Amap (AutoNavi/Alibaba Maps) CV Lab — streaming 3D reconstruction from video input only
• Key insight: uses a fixed 12-frame local context window instead of complex long-range memory mechanisms
• At each timestep: caches KV features from preceding 11 frames, predicts point map in current camera coords, estimates adjacent relative pose, recovers global trajectory via sequential pose composition
• Per-frame computation is independent of sequence length — scales to very long videos (up to 22,000 frames)
• Includes a motion-visual rotation refiner and composition-aware pose loss to limit drift
• Results: 24.45 FPS on H100, 6.71 GiB memory; ATE 4.35m on Oxford Spires (no loop closure)
• Optional loop closure module using DINOv2-SALAD for trajectory refinement on sequences with revisited regions
• Python API: simple ABotRecon.from_pretrained() → model.infer(images) interface
• Uses FlashInfer for paged KV-cache acceleration, falls back to PyTorch SDPA
• Requires: Linux, Python 3.10+, PyTorch 2.5.1, CUDA 12.1