How does DELF / DELG feature size & number impact recognition on Google Landmarks ?

Hi,

I am using DELG pre-trained on Google Landmarks v2-clean to perform landmark recognition on a large homemade dataset containing millions of images.
I adapted the extract_features.py script to generate features from the GLv2-clean dataset (instead of the given Oxford/Paris code), but I am a bit concerned about the local feature size for each image (around 500 KB per sample) since the features for the whole GLv2 dataset would take up to 1TB of space, which is more than the original dataset size.

So far, I have left the default configuration provided unchanged.
Has anyone experimented with the use_pca parameter of the config file, and how does it affect recognition/retrieval performance? I guess you could also tune the score threshold or the max number of features per image, but again, I am afraid of significatively lowering performance.

Thank you very much in advance

Start with the adapted extract_features.py workflow and the configuration controlling use_pca, score threshold, and maximum features per image. Review how these settings affect DELF/DELG feature size, then compare recognition or retrieval results on Google Landmarks v2-clean; document the observed trade-offs and recommended settings.