EgoMusic

This project contains the Python code for reproducing the results in the paper:

Alessandro Ragano, Carl Timothy Tolentino, Kata Szita, Dan Barry, Davoud Shariat Panah, Niall Murray, and Andrew Hines. 2025. EgoMusic: An Egocentric Augmented Reality Glasses Dataset for Music. In Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), October 27–31, 2025, Dublin, Ireland. ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/3746027.3758262

Setup

1. Install a recent version of ffmpeg.
2. Install the required packages by running pip install -r requirements.txt.
3. Install VISQOL by following the instructions on its GitHub page.
4. Download the EgoMusic data set by running python -m egomusic.download_egomusic. You need at least 70 GB of disk space. Alternatively, you can download the data set on Zenodo and place the files on ./data/EgoMusic/.

Audio Quality Objective Tests

1. Prepare the audio files for employing VISQOL by running python -m egomusic.prepare_visqol_files.
2. Generate the CSV file for batch processing of VISQOL by running python -m egomusic.csv_for_visqol. The CSV file should be saved in ./output/input_audio_files.csv.
3. Run VISQOL by invoking the following command on the VISQOL root directory: ./bazel-bin/visqol --batch_input_csv "path/to/egomusic/output/input_audio_files.csv" --results_csv "path/to/egomusic/output/results_audio_files.csv".
4. Obtain the objective test results by running python -m egomusic.visqol_results.

Music Source Separation

1. Download the MUSDB18-HQ data set here and save it to ./data/musdb18/. We only need the test set (50 songs) so you can delete the train folder if you want.
2. Prepare the files for separation by running python -m egomusic.process_musdb_egomusic and python -m egomusic.separation_files.
3. Employ music source separation using demucs by uploading the files in your Google Drive and running ./egomusic/source_separation.ipynb on Google Colab.
4. Download the separated files and place them on ./data/audio_for_separation/.
5. To evaluate the SDR on MUSDB and EgoMusic, run python -m egomusic.evaluate_separation.
6. To run audio quality objective tests on the stems, prepare the audio files for VISQOL analysis by running python -m egomusic.process_separated_files.
7. Generate the CSV file for batch processing by running python -m csv_for_visqol_separated. The CSV file will be saved in ./output/input_separated_files.csv.
8. Run VISQOL by invoking the following command on the VISQOL root directory: ./bazel-bin/visqol --batch_input_csv "path/to/egomusic/output/input_separated_files.csv" --results_csv "path/to/egomusic/output/results_separated_files.csv".
9. Obtain the objective test results by running python -m egomusic.visqol_results_separated.