Where to download the pose recognition model?

[chrisbin888]

Problem Description

I have reviewed the RuView repository on GitHub, which contains:

• ESP32 firmware code
• RuView Docker image

However, I could not find the download link or training method for the pose recognition model (WiFi DensePose model).

Questions

1. Where can I download the pose recognition model?
2. If I need to train it myself, where are the training data and scripts located?
3. Are there any pre-trained models available for direct use?

Environment

• ESP32-S3-DevKitC-1
• RuView Docker image

Could you please provide the model download instructions or relevant resource links? Thank you!

[d0gra]

lol

[ruvnet]

Model Availability

Thanks for the detailed question. Here's the current state:

No Pre-Trained Weights Available

RuView does not ship pre-trained pose recognition weights. This is by design — WiFi CSI-based pose estimation is highly environment-specific. A model trained in one room/building performs poorly in another because the RF propagation (multipath, reflections, attenuation) is unique to each space.

What RuView Provides Without a Trained Model

The system works in two modes:

1. Physics-based sensing (no model needed) — The ESP32 firmware + sensing server already provide:

   • Presence detection (person present/absent)
   • Person count estimation (1-3 for single node)
   • Breathing rate and heart rate estimation
   • Motion classification (absent / still / moving / active)
   • Fall detection

   This works out of the box with just firmware + server — no training required.

2. Trained model for fine-grained pose — For 17-keypoint skeleton reconstruction, you need to train a model on CSI data collected in your specific environment. This is the "DensePose" part.

How to Train

The training pipeline is built into the sensing server:

# 1. Start the sensing server
cd rust-port/wifi-densepose-rs
cargo build --release -p wifi-densepose-sensing-server
./target/release/sensing-server

# 2. Record CSI data (via the UI's "Record" button or REST API)
curl -X POST http://localhost:3030/api/recording/start?label=present_moving
# ... walk around the room for 60 seconds ...
curl -X POST http://localhost:3030/api/recording/stop

# 3. Repeat for each activity class:
#    - absent (empty room)
#    - present_still (sitting/standing still)
#    - present_moving (walking around)
#    - active (exercising, waving arms)

# 4. Train the adaptive classifier
curl -X POST http://localhost:3030/api/train

Academic Datasets (for research/benchmarking)

For the full DensePose training pipeline, the supported datasets are:

• MM-Fi — Multi-modal WiFi dataset (recommended, documented in ADR-015)
• Wi-Pose — WiFi-based pose estimation dataset

These can be used with the Rust training pipeline:

./target/release/sensing-server --pretrain --dataset /path/to/mmfi

Summary

Use Case	Model Needed?	What You Get
Presence / person count / vitals	No	Works out of the box
Activity classification (tuned to your room)	Train ~5 min of recordings	Adaptive classifier
17-keypoint skeleton	Train on MM-Fi or room recordings	Full DensePose