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

Traditional streaming 3D reconstruction systems rely on storing long-range historical data to maintain spatial consistency. As sequences grow, these systems inevitably suffer from increased latency and memory exhaustion. ABot-Recon departs from this design by operating within a fixed 12-frame local context window. By predicting only local point clouds and relative poses, the model assembles global trajectories incrementally, keeping computational complexity stable regardless of the total sequence length.

To counter the drift typically associated with local-only predictions, the developers integrated dedicated correction mechanisms that calibrate trajectory errors during both training and real-time inference. Performance data reflects these gains: on the Oxford Spires benchmark, the model reduced average trajectory error by 40.6% compared to previous methods. On the KITTI-02 dataset, it achieved a throughput of 24.45 FPS while consuming only 6.71 GB of peak memory. This efficiency allows the pipeline to run on hardware as modest as a GTX 1080 Ti.

Because the system requires only monocular RGB video input—discarding the need for depth sensors or pre-calibrated camera parameters—it offers a flexible solution for embodied AI, autonomous driving, and spontaneous 3D content production. Amap has released the inference code, pre-trained weights, and evaluation scripts on GitHub to encourage broader adoption in environments where pre-built maps are unavailable.