Amap Open-Sources ABot-Recon for Long-Sequence 3D Reconstruction

Introduction

Amap has released ABot-Recon, a streaming 3D reconstruction model designed to rebuild a scene while a camera is moving. Instead of continuously retrieving and merging an ever-growing history, the model uses only the most recent 12 consecutive frames as its local context. The goal is to keep computation from expanding with video length while supporting real-time, incremental reconstruction.

What the model changes

• No expanding long-term memory. ABot-Recon does not rely on a growing set of global memory anchors. It predicts a local point cloud in the current camera coordinate system and estimates the relative pose between neighboring frames.
• Online global assembly. Local predictions are combined as the sequence advances, gradually producing a global trajectory and a 3D scene. This is intended for “reconstruct as you capture” workflows.
• Mechanisms for drift correction. According to the project description, the model adds error constraints during training and correction mechanisms during prediction to reduce the accumulation of local pose errors.
• Minimal sensor requirements. The system takes monocular RGB video and does not require an additional depth sensor or known camera parameters. The project reports a peak memory footprint of about 6.71 GB and says that a consumer GTX 1080 Ti is sufficient to run it.

Reading the reported benchmarks

Amap reports that ABot-Recon reduced average trajectory error by 40.6% versus LingBot-Map on the Oxford Spires long-sequence benchmark, with a relative rotation error of 0.12 degrees. On KITTI-02, the reported inference speed was 24.45 FPS, or about 1.24 times that of the comparison method. The project also describes its peak memory use as roughly one-third of comparable models. These figures are reported by the project team; real-world results can vary with hardware, video input, and evaluation settings.

Why it matters

Streaming 3D reconstruction faces a basic trade-off. Long-term information can help preserve global consistency, but storing, retrieving, and fusing that information becomes increasingly expensive as a sequence grows. ABot-Recon takes the opposite route: it limits each prediction to a local window and relies on online composition and correction to maintain a usable long-range trajectory. If the approach generalizes, it could make long-video reconstruction less dependent on large memory capacity and heavy computation.

The design is potentially relevant to private-site mapping in warehouses, campuses, and malls, where positioning signals may be weak and the environment may change. It could also be explored for embodied AI, autonomous driving, and 3D content production. However, local prediction still has to cope with occlusion, rapid camera motion, weak textures, and long-term loop consistency. Benchmark numbers alone do not establish equal performance across all physical environments.

ABot-Recon has released inference and evaluation code as well as model weights on GitHub, with an experience page available on ModelScope. This gives developers a way to reproduce the reported tests and examine how the model behaves across different devices, sequence lengths, and scene types.

Source: QbitAI