HRTF Upsampling Across Varying Measurement Configurations with Geometry-Aware Query-Conditioned Aggregation
Paper โข 2609.25995 โข Published
This repository hosts the paper checkpoint and preprocessed SONICOM data for HRTF Upsampling Across Varying Measurement Configurations with Geometry-Aware Query-Conditioned Aggregation.
The implementation, dataset-preparation scripts, and training/evaluation instructions are available in the GeoAtt-HRTF GitHub repository.
| File | Description |
|---|---|
model/geoatt_final_ep440.pt |
Final GeoAtt checkpoint trained for 440 epochs on 180 listeners |
dataset/H_log.npy |
Binaural log-magnitude HRTFs in dB, shape (200, 793, 2, 106) |
dataset/locs.npy |
Source directions, shape (200, 793, 3) |
The data arrays are derived from the SONICOM HRTF dataset. Please consult and cite the original dataset when using them. The MIT license applies to the code and model, not to the underlying SONICOM data.
Clone the GitHub repository, follow its installation instructions, and run:
python tools/download_assets.py
python tools/check_data.py --hrtf-path data/sonicom/H_log.npy --directions-path data/sonicom/locs.npy
python -m vcfdconformer.evaluate \
--checkpoint weights/geoatt_final_ep440.pt \
--protocol all \
--hrtf-path data/sonicom/H_log.npy \
--directions-path data/sonicom/locs.npy \
--output-dir results/paper
@article{chen2026geoatt,
title={HRTF Upsampling Across Varying Measurement Configurations with Geometry-Aware Query-Conditioned Aggregation},
author={Chen, Xingyu and Bi, Hanwen and Zhao, Sipei and Ma, Fei and Cheng, Eva and Burnett, Ian S.},
journal={arXiv preprint arXiv:2609.25995},
year={2026}
}