Description
Convolutional Neural Networks apply learnable convolutional kernels with weight-sharing across spatial / temporal axes. They are the default backbone for CSI sensing because the natural CSI tensor (subcarriers × time × antenna pairs, or amplitude/phase spectrograms) has clear translation-invariant structure. ResNet-style residual stacks dominate the deeper variants used on crowd-density regression heatmaps.
When it's used
- Spectrogram-based HAR / gesture / gait pipelines
- Density-map regression for crowd counting
- Backbones inside
transfer-learningandself-supervised-learningsetups - Lightweight encoders for embedded CSI deployment
Limitations
- Receptive field grows slowly without dilation or pyramid pooling
- Less effective at very long temporal context vs RNN/Transformer
- Spatial assumption breaks if CSI tensor is permuted along feature axis