Amap Unveils ABot-Recon for Real-Time 3D Scene Reconstruction

Alibaba's Amap, a location-based services platform, has introduced ABot-Recon, a streaming 3D reconstruction model that can build extensive 3D scenes from a small number of frames. The technology reconstructs scenes spanning up to 10,000 frames using only 12 consecutive frames in real time, addressing limitations of conventional methods that require long-range memory. This approach reduces peak memory usage to approximately one-third of comparable systems, facilitating real-time 3D reconstruction on consumer-grade hardware.

ABot-Recon operates on a fixed 12-frame local context window, predicting local point clouds and relative poses between frames. An online composition mechanism then incrementally assembles the complete global trajectory, maintaining constant computational complexity regardless of sequence length. The model incorporates dedicated correction and constraint mechanisms to mitigate drift inherent in local predictions.

On the Oxford Spires benchmark, ABot-Recon demonstrated a 40.6% reduction in average trajectory error compared to previous leading methods, achieving a relative rotation error (RPE-R) of 0.12 degrees, which is approximately 40% lower than prior state-of-the-art. The model achieves real-time reconstruction at 24.45 FPS on KITTI-02, reportedly 1.24 times faster than existing approaches, with peak memory usage around 6.71 GB, making it feasible for use with hardware such as a GTX 1080 Ti.

The system requires only monocular RGB video as input and does not need depth sensors or pre-calibrated camera parameters. Amap positions ABot-Recon for deployment in private-area mapping, embodied AI training, autonomous driving, and 3D content production, particularly in scenarios lacking pre-built maps. The inference code, evaluation scripts, and pre-trained weights for ABot-Recon have been open-sourced on GitHub.