TALO: Pushing 3D Vision Foundation Models Towards Globally Consistent Online Reconstruction

We are still gradually removing redundant code and refining the implementation. More documentation and examples will be added soon.

Installation

1. Install TALO (Based on VGGT-SLAM)

sudo apt-get install git python3-pip libboost-all-dev cmake gcc g++ unzip   # required by VGGT-SLAM
git clone https://github.com/TODO/talo
cd talo
conda create -n talo python=3.11
conda activate talo
./setup.sh   # installs third-party dependencies required by VGGT-SLAM 

2. Install Additional 3D Vision Foundation Models (Optional)

TALO currently supports the following 3D Vision Foundation Models as interchangeable backbones.

It is also easy to integrate more advanced 3DVFMs by only formatting the prediction as a python dictionary containing the following keys (see VFMs_adaptor.py for example implementations):

• "org_images"
• "images"
• "cam2world"
• "intrinsic"
• "world_points"
• "world_points_conf"

Dataset Preparation

TALO currently supports:

• nuScenes
• Waymo Open Dataset

1. Download Raw Datasets

Waymo Open Dataset

Download raw .tfrecord sequences: https://waymo.com/open/

nuScenes

Download the full dataset: https://www.nuscenes.org/download

2. Convert Raw Data into TALO Format

After downloading, convert datasets using our extraction scripts:

python dataset/extract_waymo.py
python dataset/extract_nuscenes.py

Note that parsing Waymo requires waymo-open-dataset-tf-2-6-0, which depends on older package versions (e.g., NumPy 1.x) and is not compatible with the TALO environment. Therefore, please create a separate Python environment specifically for extracting Waymo.

Please modify data_root and save_root accordingly in each script.

These scripts will:

• Extract RGB images (as model input)
• Extract camera intrinsics, extrinsics, and LiDAR (used as GT for evaluation)

Output directory format

scene_dir/
  image/
    FRONT/
      000.jpg
      ...
  cam2world/
    FRONT/
      000.txt        # 4x4 matrix
      ...
    ...
  intrinsic/
    FRONT.txt        # 3x3 matrix
    ...
  lidar/
    000.bin
    ...

Custom Data

To run the system on your own data, format it as follows:

custom_data
  example_scene/
    image/
      cam0/
        000.jpg
        ...
      ...

Then run:

python main.py --data_folder ./Data/custom_data/example_scene/ --log_path ./Save/custom_data/example_scene/VGGT+60+tps

for reconstruction and

python eval_vis_pcd_traj.py --GT ./Data/custom_data/example_scene/ --pred ./Save/custom_data/example_scene/VGGT+60+tps --vis

for visualization.

Run and Evaluation

We provide a quick-start script that runs TALO on both Waymo and nuScenes, and summarizes results as reported in the paper.

bash run.sh

main.py — Key Arguments

Visualization

TALO provides both online and offline visualization modes.

1. Incremental Online Visualization (viser)

Online VGGT-SLAM visualization by adding:

--vis_map

to main.py (inside run.sh).

2. Offline Full Visualization (Open3D)

Enable offline reconstruction visualization by adding:

--vis

to eval_vis_pcd_traj.py (inside run.sh).

Examples

VGGT on Waymo

VGGT on nuScenes

Pi3 on Waymo

Pi3 on nuScenes

Acknowledgements

To ensure fair comparisons between different submap alignment methods (SL4 from VGGT-SLAM and Sim3 from VGGT-Long), TALO is built upon the same framework (VGGT-SLAM) and extended to support multi-camera settings as well as additional 3D Vision Foundation Models (3DVFMs), including VGGT, Pi3, and MapAnything. All rights of these projects are fully reserved by their respective authors.

We sincerely thank the authors and maintainers of these outstanding open-source projects. If you find TALO useful, please consider citing and starring our work, and supporting the projects that made it possible.