🎯 ECA Presenter - ECAGestureNet

On-Device AI Remote for Slide Control Using Hand Gestures

ECA Presenter is a lightweight on-device AI remote that lets you control your presentation slides using webcam-based hand gesture recognition — no Bluetooth, smartphone, or network required.

Built on ECA-Net (Efficient Channel Attention), it performs real-time gesture inference on CPU-only environments at up to 30 FPS.

---

✋ Supported Gestures

Gesture	Action	Model Label	Description
✋ Palm	Next slide	fist	Palm and fist are unified under the same label (fist) and mapped to “Next Slide.”
👌 OK Sign	Previous slide	ok	Thumb and index finger form a circle.
👉 Index Up	Activate laser pointer	index_up	Triggers the pointer shortcut (e.g., Ctrl + L in PowerPoint).
✌ V Sign	End presentation	v_sign	Ends the presentation and disables the pointer.

Compatible with PowerPoint, Keynote, and Google Slides.

---

ECAGestureNet Architecture

──────────────────────────────────────────────────────────────
Input: 3 × 224 × 224 RGB
──────────────────────────────────────────────────────────────
Stage 1: Conv(3→32, k3, s2, p1) → BN → ReLU → ECA(32)
Output: 32 × 112 × 112
──────────────────────────────────────────────────────────────
Stage 2: Conv(32→64) → BN → ReLU → ECA(64)
Output: 64 × 56 × 56
──────────────────────────────────────────────────────────────
Stage 3: Conv(64→128) → BN → ReLU → ECA(128)
Output: 128 × 28 × 28
──────────────────────────────────────────────────────────────
Stage 4: Conv(128→256) → BN → ReLU → ECA(256)
Output: 256 × 14 × 14
──────────────────────────────────────────────────────────────
Global AvgPool → FC(256 → num_classes)
Output: logits (4 classes)
──────────────────────────────────────────────────────────────

Highlights

• Each stage uses Conv-BN-ReLU + ECA Block
• ECA (Efficient Channel Attention) applies 1D convolution-based channel attention
• Lightweight alternative to SE/CBAM with minimal overhead
• Global Average Pooling + FC for classification (ok, fist, index_up, v_sign)

Summary: “Four Conv-ECA stages + Global Pool + FC” = compact yet powerful gesture recognition CNN.

---

How to Use

1) Clone the repository

git clone https://github.com/USER/eca_presenter.git
cd eca_presenter

2) Create a virtual environment

python -m venv .venv
# Windows
.\.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

3) Install PyTorch

Use the official PyTorch installer for your system:
🔗 https://pytorch.org/get-started/locally/

Example (CPU only):

pip install torch torchvision

4) Install remaining dependencies

pip install -r requirements.txt

---

2. Run the Application

python runtime/main.py

Runtime behavior:

• Displays the recognized gesture and confidence score.
• Sends keyboard events directly to the active presentation window.
• Works fully offline using ONNX Runtime.

Included models:

models/gesture_eca.onnx
assets/labels.txt

---

3. Training & Model Conversion

Dataset structure

data/
  train/
    ok/
    fist/
    index_up/
    v_sign/
  val/
    ok/
    fist/
    index_up/
    v_sign/

Train the model

python model/train_eca_gesture.py

Output:

model/eca_gesture.pth
assets/labels.txt

Convert to ONNX

python model/export_onnx.py

Output:

models/gesture_eca.onnx

---

4. Project Structure

eca_presenter/
├── model/
│   ├── train_eca_gesture.py        # Training script
│   └── export_onnx.py              # ONNX exporter
├── runtime/
│   └── main.py                     # Webcam runtime + slide control
├── models/
│   └── gesture_eca.onnx            # Trained ONNX model
├── assets/
│   └── labels.txt                  # Class labels
├── requirements.txt
└── README.md

---

5. Design Motivation

① Inefficiency of smartphone remotes

Presenters often can’t use both hands freely during talks.
Using a smartphone to swipe slides interrupts the flow.

② Limitations of Bluetooth clickers

• Battery drain or pairing failure
• Compatibility issues
• Easy to lose
• May disconnect unexpectedly

③ On-device AI advantages

• No internet required
• No data sent externally (privacy-safe)
• Runs in real time on CPU using ONNX Runtime
• Minimal latency and stable slide control

---

6. System Pipeline

1. MediaPipe Hands detects the hand region.
2. Crop and resize ROI to 224×224.
3. ONNX Runtime performs gesture inference via ECAGestureNet.
4. Apply stability filtering (confidence & consistent frames).
5. Send key events using pyautogui/keyboard to control slides.

Achieves ~30 FPS on CPU with < 50 ms end-to-end latency.

---

7. Research Contribution

Goal	Description
ECA validation in real HCI	Demonstrates ECA’s effectiveness in real-time, on-device gesture recognition.
Lightweight attention	Achieves similar accuracy to SE/CBAM with fewer FLOPs.
Realtime performance	Runs on CPU with no perceptible delay.
Applied prototype	Integrates ECA-Net into a functional presentation-control application.

This project bridges academic model design and practical on-device AI applications in HCI.

---

8. Development Environment

• Python 3.10
• PyTorch / ONNX / ONNX Runtime
• OpenCV
• MediaPipe (optional)
• keyboard / pyautogui

---

9. Example Use Cases

• Gesture-controlled slide navigation during live talks
• Online teaching with natural pointer control
• Interactive media art installations
• Conference rooms without physical remotes

---

Reference

Wang Q., Wu B., Zhu P., Li P., Zuo W., Hu Q.
ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks.
Proceedings of CVPR 2020.