CSI to Depth (generative)

[Arritmic]

Missing a powerful feature for downstream tasks based on well established computer vision pipelines:

https://github.com/Multimodal-Sensing-Lab/csi2depth

[ruvnet]

Great reference — the CSI2Depth work from Multimodal-Sensing-Lab is directly relevant. Their approach uses a generative model to convert CSI amplitude maps into depth images, which could complement our current pipeline.

Current RuView architecture and where CSI2Depth fits:

Layer	Current	With CSI2Depth
CSI capture	ESP32 → UDP → server	Same
Feature extraction	Variance, spectral power, subcarrier weights	CSI amplitude map as 2D input
Pose estimation	Heuristic skeleton from features	Depth map → standard depth-based pose
Person separation	DynamicMinCut on correlation graph	Depth segmentation

Integration path:

1. Collect CSI amplitude matrices (already available in `frame_history`)
2. Train the CSI2Depth generative model on paired CSI + depth data
3. Feed generated depth maps to standard pose estimators (e.g., MoveNet, MediaPipe)

This would be a significant quality jump — going from WiFi-derived features to actual depth images. Tagged as enhancement for future work.

[Arritmic]

Btw, a video can be see here:
https://www.youtube.com/watch?v=tNGM4l1PZQQ