Technical Inquiry: Spatial Resolution and Phase Coherence Constraints in ESP32-based DensePose

[bin9208]

"I acknowledge that ESP32 CSI is capable of coarse sensing, such as respiration, heart rate, and occupancy detection, as these rely on temporal amplitude fluctuations. However, DensePose requires high spatial resolution and phase coherence."

"How do you overcome the 'Spatial Rank Deficiency' of a SISO device? Even with a mesh of 3 ESP32s, the lack of a shared Local Oscillator (LO) means you cannot perform coherent beamforming or AoA estimation. Are you essentially claiming that your model can 'hallucinate' 3D body surfaces from asynchronous, 1D amplitude data? This seems more like a statistical guess than actual sensing."

[ruvnet]

This is a fair and important question. Let me be direct about what the system actually does and where the limitations are.

What RuView Actually Senses

You're correct that ESP32 CSI from a SISO device cannot provide the spatial resolution needed for true DensePose (UV surface mapping). Here's an honest breakdown:

What Works Reliably (Physics-Based)

Capability	Mechanism	Accuracy
Presence detection	Amplitude variance exceeds baseline	High (>95%)
Person count (1-3)	Subcarrier variance clustering	Moderate (with multiple nodes)
Breathing rate	0.1-0.5 Hz bandpass on phase	Good (±2 BPM in controlled settings)
Heart rate	0.8-2.0 Hz bandpass on phase	Moderate (requires stillness)
Motion classification	Temporal variance + spectral energy	Good
Fall detection	Phase acceleration threshold	Moderate

These rely on temporal amplitude/phase fluctuations, as you noted — not spatial resolution.

What is Approximate / Aspirational

Capability	Reality
17-keypoint skeleton	Derived from CSI features (variance, motion energy, breathing amplitude) mapped to plausible poses — not reconstructed from spatial data. You're right to call this closer to a "statistical estimate" than true sensing.
DensePose UV mapping	Not achievable with SISO ESP32. The name references the CMU research direction, not a claim of equivalent results.
3D body surface	Not possible without coherent antenna arrays or camera ground truth.

The Multi-Node Advantage

With a mesh of 3+ ESP32 nodes, you get:

• Spatial diversity — each node sees a different multipath profile, so cross-node variance provides coarse spatial information
• But NOT coherent beamforming — you're correct that without a shared LO, you can't do phase-coherent AoA. Each node's phase is relative to its own oscillator.
• The system uses amplitude-based features across nodes (per-node state since PR feat(server): per-node state pipeline for multi-node sensing (#249) #326), not phase coherence between nodes.

What Makes This Useful Despite Limitations

1. Privacy — no cameras, no images to leak
2. Cost — $9/node vs. $50+ for mmWave radar
3. Coverage — WiFi signals pass through walls
4. The sensing floor is real — presence, occupancy, vital signs, and activity classification are genuinely useful for elder care, smart buildings, and security — even without full DensePose.

The Alpha Notice

We updated the README to reflect this honestly:

Alpha Software — Multi-node person counting may show identical output (#249). No pre-trained model weights provided. Training from scratch required.

The project is research-grade software exploring what's possible with commodity WiFi hardware. The skeleton visualization is a rendering of the classified activity state, not a reconstruction from spatial data. We should be clearer about this in the documentation, and your feedback helps.

References:

• ADR-068: Per-node state pipeline (architectural improvement for multi-node)
• ADR-024: Contrastive CSI embedding (AETHER — aspirational learned embeddings)
• CMU DensePose From WiFi paper (the research inspiration, not a claim of equivalence)