Feature extraction and save them from all the images and point-cloud pairs?

[shahriar0651]

How do we use the pre-trained unimodal feature extractors and save the feature representation pairs from all the images and point cloud pairs? The end goal is not to deploy a detection head; instead, use an intermodality consistency check on the feature representation (using representation learning) for anomaly detection. Any directions for that? Thanks!