Amap Releases ABot-Recon for Streaming 3D Reconstruction

Amap, Alibaba Group's location based services platform, announced in a press release ABot-Recon, a streaming 3D reconstruction model that uses 12 consecutive frames to reconstruct scenes spanning more than 10,000 frames in real time.

The model works without long range memory anchors. It predicts a local point cloud and the relative pose between adjacent frames, then uses an online composition process to build the full global trajectory over time.

Amap said ABot-Recon achieved 24.45 FPS on KITTI-02 with peak memory use of about 6.71 GB, allowing the full pipeline to run on a GTX 1080 Ti. On the Oxford Spires long sequence benchmark, it reduced average trajectory error by 40.6 percent compared with the previous leading method.

ABot-Recon uses monocular RGB video as input and does not require depth sensors or pre calibrated camera parameters. Amap has released the inference code, evaluation scripts, and pre trained weights on GitHub.