amap-cvlab/ABot-Recon

What it solves

ABot-Recon addresses the challenge of long-horizon streaming 3D reconstruction. Traditional methods often rely on complex mechanisms to maintain long-range state or memory, which can become computationally expensive as the sequence length increases. ABot-Recon simplifies this by focusing on local context to maintain efficiency and constant memory usage regardless of the sequence length.

How it works

Instead of persistent long-range memory, the system uses a fixed 12-frame local context window. At each time step, it performs the following:

1. KV-Cache: It caches Key-Value (KV) features from the preceding 11 frames.
2. Geometry Prediction: It predicts a point map in the current camera coordinate system.
3. Pose Estimation: It estimates the adjacent relative pose between the current and previous frame.
4. Global Composition: It recovers the global trajectory and point cloud by sequentially composing these relative poses.

To minimize drift over long sequences, it employs a lightweight motion-visual rotation refiner and a composition-aware pose loss. An optional loop-closure backend (using DINOv2-SALAD descriptors) can be used to further refine the trajectory when the camera revisits known regions.

Who it’s for

Researchers and developers working in computer vision, robotics, and 3D reconstruction who need a memory-efficient, streaming-capable approach to mapping and trajectory estimation from video streams.

Highlights

• Constant Memory Footprint: Per-frame computation and model-state memory remain independent of the sequence length.
• High Efficiency: Achieves up to 24.45 FPS on an NVIDIA H100.
• Local-First Approach: Reconstructs long video streams using only a 12-frame local context window.
• Optional Loop Closure: Supports trajectory refinement via sparse pose-graph optimization for revisited regions.