Amap's ABot-Recon Rebuilds 3D Worlds From Live Video Using Under 7GB
Amap CV Lab open-sourced ABot-Recon, a streaming 3D reconstruction model that runs on a rolling 12-frame window at 24 FPS.
PRO- Amap CV Lab open-sourced ABot-Recon, streaming 3D reconstruction from monocular video under Apache 2.0.
- Uses a fixed 12-frame rolling window, so memory and compute stay constant regardless of video length.
- Hits 24.45 FPS at 6.71 GiB GPU memory on KITTI-02, H100, 504x280 resolution.
- Oxford Spires: 4.35 m ATE, 0.12 degrees RPE-R, and 91.81 percent F1 dense reconstruction without loop closure.
- Optional loop-closure backend uses DINOv2-SALAD retrieval plus pose-graph optimization for revisited scenes.
- Weights on Hugging Face; training code slated for later release.
Reconstructing a 3D world from video usually means loading long sequences into VRAM and running slow offline optimization. ABot-Recon, a newly open-sourced model from Amap CV Lab, processes video as a live stream using a fixed 12-frame local context, reconstructing both the camera trajectory and a global point cloud on the fly. The code, weights, and a technical report are all available under Apache 2.0.
At each timestep, the model caches KV features from the preceding 11 frames, predicts a point map in the current camera coordinate system, estimates the adjacent relative pose between the previous and current frame, and recovers the global trajectory and point cloud through sequential pose composition. Because the window is fixed, both model-state memory and per-frame computation stay constant regardless of sequence length.
Why a fixed window works
Most long-horizon SLAM and neural reconstruction pipelines lean on increasingly elaborate mechanisms to retain and fuse long-range state. ABot-Recon sidesteps that entirely. At every timestep it solves the same bounded prediction problem regardless of how long the video has been running.
The obvious risk is drift: compose enough small pose errors and the global trajectory bends. To limit that, the team adds a lightweight motion-visual rotation refiner and a composition-aware pose loss so that local poses hold up when chained over long horizons.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
