Amap releases ABot-Recon for long-video 3D reconstruction

Twelve frames at a time

Amap's ABot-Recon reconstructs three-dimensional scenes from a stream of video frames while retaining a fixed local context. The important qualification is what its 12-frame claim means: the system uses the current frame together with cached information from the preceding 11. It continues to process new frames as the video proceeds.

It does not infer an entire long recording from only 12 photographs. A sliding window limits the model's active memory, while the accumulated camera trajectory and reconstructed scene can extend across a much longer sequence. This distinction changes the meaning of the original announcement from an extreme sparse-input claim to an efficiency result for streaming reconstruction.

How local estimates become a map

The August 27 technical paper describes two predictions at each step: a map of points in the current camera's coordinate system and the camera's motion relative to the preceding frame. Successive relative motions are composed to place those local estimates into a common global space.

Small orientation mistakes can accumulate as that chain grows. The researchers address this with a lightweight rotation-refinement component using recent visual and motion information, plus a training loss that evaluates composed poses across multiple steps. The method therefore targets the drift created by repeatedly joining local estimates, rather than retaining an ever-growing learned memory of the full video.

The conditions behind the results

The project reports an absolute trajectory error of 4.35 meters and rotational relative-pose error of 0.12 degrees on Oxford Spires without loop closure. Its repository also lists 24.45 frames per second and 6.71 GiB of memory for KITTI-02 at 504 by 280 resolution on an Nvidia H100. Input storage is excluded from that memory figure.

These measurements refer to different properties. Trajectory error measures the estimated camera path; processing speed describes throughput under a specified hardware and image-size setting. A constant model-state footprint does not mean that saving every output point cloud, original frame and visualization consumes a constant amount of disk space.

The reproducible software environment

The release targets Linux with Python 3.10 or newer, PyTorch 2.5.1 and CUDA 12.1. The maintainers distinguish the A100 machine used to validate installation from the H100 used for the quoted runtime measurement. That prevents a reader from treating every listed accelerator as the source of the 24.45-frame-per-second result. FlashInfer supplies accelerated cache operations where available, with a PyTorch fallback; an additional compiled rotary-position component is optional. These software choices are part of the setup to record when comparing local performance with the paper.

Released tools and an optional refinement stage

The public package includes checkpoints, a Python interface and a demonstration that exports camera poses, local point maps and confidence information. An optional loop-closure stage can refine a trajectory when a recording revisits the same location; the learned streaming model itself does not depend on it.

The repository separately publishes code and model license documents. It also records a September community port to Axera's AX650N processor. That port is a subsequent hardware adaptation, not the hardware used for the headline H100 performance measurement. Together, the release and its documented evaluation make ABot-Recon a reproducible approach to bounded-memory mapping, while preserving the distinction between the core streaming method and additional optimization.

Axera's own model card reports approximately three seconds per frame for its port, with about 5.3 GiB of reserved memory. It also documents the absence of loop closure and a reduced effective cache setting as current limitations. The port's weights retain non-commercial CC BY-NC 4.0 terms, while the conversion and inference code is Apache 2.0. These details make clear that a hardware adaptation can have different speed, behavior and licensing boundaries from the original research setup.