Description
Data augmentation produces synthetic training samples by applying label-preserving transformations to real ones. CSI augmentation is its own subdomain — time warping, subcarrier dropout, antenna permutation, additive noise, simulated multipath — designed to enlarge a small CSI corpus without collecting more data. It is the most cost-effective regulariser for cross-domain CSI sensing.
When it's used
- Bridging small labelled CSI sets to deployment scale
- Robustness training against hardware noise
- Pretext task for
self-supervised-learning(contrastive pairs come from augmentations)
Limitations
- Bad augmentations can destroy labels (e.g. Doppler-distorting time warp on gait)
- Domain-specific design — transferring vision augmentations to CSI is non-trivial
- Marginal gains shrink as labelled data grows