This is the official pyTorch implementation of the CVPR24 paper "The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement". In this work, we present a simple approach for Pose Refinement that combines pre-trained features with a particle filter and a renderable representation of the scene.
[CVPR 2024 Open Access] [ArXiv]
Our proposed Pose Refinement algorithm.
The following command
$ python download.py
Will download colmap models, and pre-trained Splatting models for the scene representation, as well as the Cambridge Landmark dataset.
Follow the instructions to install the gaussian_splatting environment from the official repo
Then, activate the environment and execute:
$ pin install -r requirements.txt
To reproduce results on Cambridge Landmarks, e.g. for KingsCollege:
$ python refine_pose.py KingsCollege --exp_name kings_college_refine --renderer g_splatting --clean_logs
The script will load the config for the number of steps and hyperparameters of the MonteCarlo optimization from the configs.py file. It will utilize a Gaussian Splatting model to render candidate poses. Other options such as a colored mesh, or a NeRF model will be uploaded soon
Coming soon
Coming soon
Here is the bibtex to cite our paper
title={The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement},
author={Trivigno, Gabriele and Masone, Carlo and Caputo, Barbara and Sattler, Torsten},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={12786--12798},
year={2024}
}
Parts of this repo are inspired by the following repositories: