ActLoc:
Learning to Localize on the Move via Active Viewpoint Selection

Jiajie Li* · Boyang Sun* · Luca Di Giammarino · Hermann Blum · Marc Pollefeys

CoRL 2025
Paper | Webpage | Interactive Demo |

We present ActLoc, a learning-based approach for active viewpoint selection. ActLoc selects viewing directions at given locations that maximize visual localization performance.

🔄 Updates

• Sep 22, 2025: Check out our online interactive demo!
• Sep 21, 2025: Initial code release

---

🛠️ Setup

Our code has been tested on a workstation running Ubuntu 22.04 with an NVIDIA RTX 4090 (24GB) GPU and CUDA 12.4.

We recommend creating the environment from the provided environment.yml file:

# create environment
conda env create -f environment.yml

# activate environment
conda activate actloc_env

---

Example Data Download

We provide one example scene from HM3D:

chmod +x download_example_data.sh && ./download_example_data.sh

You can also download the example data from here and put it in the root folder of this repo to use for demos.

---

Quick Start

Single Viewpoint Selection

This script takes an SfM model and a list of waypoints as input, and predicts the optimal camera pose (extrinsic matrix) for each one individually.

python demo_single.py \
    --sfm-dir ./example_data/reference_sfm \
    --waypoints-file ./example_data/waypoints.txt \
    --checkpoint ./checkpoints/trained_actloc.pth \
    --output-file ./example_data/selected_poses.npz

Expected Output: This will create a file named selected_poses.npz in the example_data directory, which contains the calculated poses.

Multiple Viewpoints Selection with Motion Constraints

To choose viewpoint that balances localization performance and rotation continuity, run:

python demo_multi.py \
    --sfm-dir ./example_data/reference_sfm \
    --waypoints-file ./example_data/waypoints.txt \
    --checkpoint ./checkpoints/trained_actloc.pth \
    --output-file ./example_data/selected_poses_multi.npz \
    --lamda 0.02

Visualization

To visualize the scene and the prediction results:

python vis.py \
    --meshfile ./example_data/scene.glb \
    --waypoints-file ./example_data/waypoints.txt \
    --poses-file ./example_data/selected_poses.npz 

or:

python vis.py \
    --meshfile ./example_data/scene.glb \
    --waypoints-file ./example_data/waypoints.txt \
    --poses-file ./example_data/selected_poses_multi.npz 

Render Images from Predicted Poses

To see the observation ath selected poses, use the selected_poses.npz file generated to render images from the 3D mesh.

python capture_predicted_views.py \
    --mesh-file ./example_data/scene.glb \
    --poses-file ./example_data/selected_poses.npz \
    --output-folder ./example_data/best_viewpoint_images

Expected Output: This will create a new folder named best_viewpoint_images containing the rendered images for each successful waypoint. This process might take a while.

---

📝 Roadmap

• Single-viewpoint Selection Inference Code Release
• Single-viewpoint Selection Demo Release
• Path Planning Inference Code Release
• Path Planning Demo Release
• Test Data and Evaluation Script Release
• Training Data and Training Script Release

---

Citation

If you find our paper useful, please consider citing:

@misc{li2025actloclearninglocalizeactive,
      title={ActLoc: Learning to Localize on the Move via Active Viewpoint Selection}, 
      author={Jiajie Li and Boyang Sun and Luca Di Giammarino and Hermann Blum and Marc Pollefeys},
      year={2025},
      eprint={2508.20981},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2508.20981}, 
}