Skip to content

This is the official implementation of the paper "Rotation Synchronization Via Deep Matrix Factorization" ICRA 2023

Notifications You must be signed in to change notification settings

gktejus/DMF-synch

Repository files navigation

Rotation Synchronization via Deep Matrix Factorization

GK Tejus, Giacomo Zara, Paolo Rota, Andrea Fusiello, Elisa Ricci, Federica Arrigoni

This is the official Code for the paper Rotation Synchronization via Deep Matrix Factorization presented at ICRA 2023.

Setup

Our current implementation is based off Implicit Regularization in Deep Matrix Factorization . Please use python 3.7 to run this code.

The basic libraries are listed in requirements.txt. You can install them by running the following command.

pip install -r requirements.txt

Another way to do this is by using the setup bash script. This will install the required libraries and download and arrange the datasets. This can be done by running the following command.

bash setup.sh

Datasets and Configs ( 1DSfM / Synthetic )

1DSfM

For our experiments, we use the 1DSfM dataset and synthetic datasets. The 1DSfM dataset can be downloaded from here.

The general structure of the datasets and configs for a particular dataset is as follows.

.
└── DMF-Synch
    ├── datasets
    │   ├── data-real
    │   │   ├── [dataset]_gt.pt ------> Ground truth relative rotations (noisy)
    │   │   ├── [dataset]_obs.pt -----> Observed indices
    │   │   └── [dataset]_unobs.pt ---> Unobserved indices
    │   └── data-real-mat
    │       └── [dataset].mat --------> Ground truth absolute rotations
    └── configs
        └── data-real
            └── [dataset].toml --------> Config file

The easiest way to setup the 1DSfM dataset is to run the bash script setup.sh. Please note that this will download and arrange all the 15 datasets contained in the 1DSfM dataset. The configs for each dataset in 1DSfM are already present in the configs folder.

Synthetic Datasets

To generate the synthetic datasets, you'll first have to generate the ground truths and the mask using the MATLAB script synth_real.m located in the MATLAB-scripts folder.

This script has several paramerters that can be changed. The parameters are as follows:

  • ncams_c : Number of cameras
  • prob_out : Fraction of outliers
  • sigma_a : Noise in degrees
  • fraction_missing : Fraction of missing data

After setting the parameters and running the MATLAB script, to generate the final datasets and configs, please run the following command.

python gen_synth.py 

Experiments

If you just want to run one experiment, use the following command as an example.

python3 main.py --print_config --log_dir ./logs \
    --config configs/mat-real/Alamo.toml, \
    --config configs/opt/SGD.toml, \
    --set depth 2, \
    --set lr 0.001, \
    --set initialization gaussian, \
    --set init_scale 1.e-3

To run multiple experiments with multiple hyperparameters sequentially, you can use ./scripts/run.rb (please make sure Ruby is installed and gem install colorize --user). The code will log into ~/logs by default.

./scripts/run.rb --n_jobs 1 --name mat-cmpl --template 'python3 main.py \
    --print_config --log_dir ./logs/ --config configs/data-real/CONFIG, \
    --config configs/opt/SGD.toml, --set depth DEPTH --set lr LR \
    --set init_scale SCALE  --set n_iters 100000 --set initialization INIT'\
    --replace LR=0.3 \
    --replace DEPTH=2,3 \
    --replace SCALE=1.e-3 \
    --replace CONFIG=Alamo.toml, Ellis_Island.toml\ 
    --replace INIT=gaussian

Citation

If you found our work useful and use it in your research, please consider citing our paper.

@misc{tejus2023rotation,
  title={Rotation Synchronization via Deep Matrix Factorization}, 
  author={Gk Tejus and Giacomo Zara and Paolo Rota and Andrea Fusiello and Elisa Ricci and Federica Arrigoni},
  year={2023},
  eprint={2305.05268},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

If you have any questions regarding the code, please raise an issue or reach out to me.

About

This is the official implementation of the paper "Rotation Synchronization Via Deep Matrix Factorization" ICRA 2023

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published