SCRStudio is a unified and modular framework for Scene Coordinate Regression (SCR)-based visual localization, built on top of the nerfstudio project.

This library provides an interpretable and modular implementation of SCRs, breaking down components such as input encoding, network architecture, and supervision strategies. It offers a unified implementation of three major SCR methods: ACE, GLACE, and R-SCoRe. SCRStudio supports various pretrained local encodings (both sparse and dense) while incorporating state-of-the-art techniques for integrating global encodings.

This guide will help you get started with the default R-SCoRe SCR model trained on the classic Aachen dataset.

We recommend using conda to manage dependencies. Make sure to install Conda before proceeding.

Install PyTorch with CUDA (tested with CUDA 12.1 and 12.4). PyTorch Geometric and cuML are also required for encoding preprocessing.

For CUDA 12.4:

conda create -n scrstudio python=3.10 pytorch=2.5.1 torchvision=0.20.1 pytorch-cuda=12.4 cuml=25.02 -c pytorch  -c rapidsai -c conda-forge -c nvidia
conda activate scrstudio
pip install --upgrade pip
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.5.1+cu124.html

git clone --recursive https://github.com/cvg/scrstudio.git
cd scrstudio
pip install --upgrade pip setuptools
pip install -e .

The following steps will train a scrfacto model, our recommended model for large scenes.

# Download Aachen dataset:
scr-download-data aachen
# Download specific capture for NAVER Lab dataset:
scr-download-data naver --capture-name dept_1F

scrfacto follows the methodology from R-SCoRe (paper), utilizing PCA for local encoding dimensionality reduction and Node2Vec for learning global encodings.

To reduce GPU memory usage for local encoding buffer, apply PCA on local encodings:

# Compute PCA for Dedode local encodings on Aachen dataset
scr-encoding-pca dedode --encoder.detector L --encoder.descriptor B --n_components 128 --data data/aachen

This saves the PCA components as PyTorch state dictionary at: data/aachen/proc/pcad3LB_128.pth

Compute the pose overlap score for training images:

scr-overlap-score --data data/aachen/train --max_depth 50

This saves a sparse COO format overlap matrix at: data/aachen/train/pose_overlap.npz

Train a Node2Vec model on this graph:

scr-train node2vec --data data/aachen --pipeline.model.graph pose_overlap.npz --pipeline.model.edge_threshold 0.2

Use the trained global encoding:

cp outputs/aachen/node2vec/<timestamp>/scrstudio_models/head.pt data/aachen/train/pose_n2c.pt

Now, train the scrfacto model:

scr-train scrfacto --data data/aachen --pipeline.datamanager.train_dataset.feat_name pose_n2c.pt

Results are saved in outputs/aachen/scrfacto/<timestamp>.

Compute NetVLAD retrieve features and compress them with Product Quantization (PQ):

scr-retrieval-feat --data data/aachen/train --pq

Results are saved in data/aachen/train/netvlad_feats_pq.pkl

Compute retrieval features for test images and run evaluation:

scr-retrieval-feat --data data/aachen/test
scr-eval --load-config outputs/aachen/scrfacto/<timestamp>/config.yml --split test

This code builds on previous camera relocalization pipelines, namely DSAC, DSAC++, DSAC*, ACE, GLACE, and R-SCoRe. Please consider citing:

@inproceedings{brachmann2017dsac,
  title={{DSAC}-{Differentiable RANSAC} for Camera Localization},
  author={Brachmann, Eric and Krull, Alexander and Nowozin, Sebastian and Shotton, Jamie and Michel, Frank and Gumhold, Stefan and Rother, Carsten},
  booktitle={CVPR},
  year={2017}
}

@inproceedings{brachmann2018lessmore,
  title={Learning less is more - {6D} camera localization via {3D} surface regression},
  author={Brachmann, Eric and Rother, Carsten},
  booktitle={CVPR},
  year={2018}
}

@article{brachmann2021dsacstar,
  title={Visual Camera Re-Localization from {RGB} and {RGB-D} Images Using {DSAC}},
  author={Brachmann, Eric and Rother, Carsten},
  journal={TPAMI},
  year={2021}
}

@inproceedings{brachmann2023ace,
    title={Accelerated Coordinate Encoding: Learning to Relocalize in Minutes using RGB and Poses},
    author={Brachmann, Eric and Cavallari, Tommaso and Prisacariu, Victor Adrian},
    booktitle={CVPR},
    year={2023},
}

@inproceedings{wang2024glace,
    title={Glace: Global local accelerated coordinate encoding},
    author={Wang, Fangjinhua and Jiang, Xudong and Galliani, Silvano and Vogel, Christoph and Pollefeys, Marc},
    booktitle={CVPR},
    year={2024}
}

@inproceedings{jiang2025rscore,
      title={R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual Localization},
      author={Jiang, Xudong and Wang, Fangjinhua and Galliani, Silvano and Vogel, Christoph and Pollefeys, Marc},
      booktitle = {CVPR},
      year={2025}
}