Tsinghua-MARS-Lab/SLAM-Former

[ECCV 2026] SLAM-Former: Putting SLAM into One Transformer

What it solves

SLAM-Former addresses the complexity of Simultaneous Localization and Mapping (SLAM) by unifying the entire process into a single Transformer architecture. It aims to provide a streamlined approach to 3D reconstruction and camera tracking from image sequences.

How it works

The project implements a Transformer-based model that processes image sequences to perform SLAM. It includes features like KV pruning to optimize inference and a ConvHead variant to reduce grid artifacts in the output. The system supports both real-time interactive visualization via Rerun and static browser-based visualization of the final 3D point clouds and trajectories.

Who it’s for

This tool is designed for researchers and developers working in computer vision, robotics, and 3D scene reconstruction who want to leverage Transformer-based architectures for spatial AI.

Highlights

  • Unified Architecture: Integrates the SLAM process into one Transformer model.
  • Efficient Inference: Supports KV pruning to manage memory and computation during sequence processing.
  • Flexible Visualization: Offers both real-time interactive 3D viewing and static HTTP-based result visualization.
  • Diverse Training Data: Trained on multiple large-scale datasets including ARKitScenes, ScanNet, and MegaDepth.

Related

  • Project
    MIT-SPARK/VGGT-SLAM

    A real-time dense feed-forward scene reconstruction system for 3D mapping and localization, featuring optional open-set 3D object detection.

  • Project
    PeterFWS/Structure-PLP-SLAM

    A visual SLAM system that uses points, lines, and planes for efficient sparse mapping and localization across monocular, RGB-D, and stereo cameras.

  • Project
    transformerlab/transformerlab-app

    An open-source machine learning platform that unifies training, fine-tuning, inference, and evaluation into a single interface for individuals and research labs.

  • Project
    sjtuyinjie/Ultra-Fusion

    A resilient, tightly-coupled multi-sensor SLAM framework that maintains accurate localization for wheeled, legged, and aerial robots even during sensor degradation or calibration errors.

  • Project
    luigifreda/pyslam

    A hybrid Python/C++ modular framework for Visual SLAM that integrates modern AI depth prediction, semantic segmentation, and volumetric reconstruction for research and prototyping.