YOLO Experiment

A collection of Python scripts for experimenting with YOLO (You Only Look Once) models for various computer vision tasks including object detection and pose estimation.

Overview

This repository contains implementations for running YOLO models on different input sources (images, webcam, phone camera) with support for object detection, pose estimation, and instance segmentation tasks. Includes both command-line scripts and an interactive Streamlit web application for easy experimentation.

Repository Structure

yolo-experiment/
├── app.py               # Streamlit web application
├── object-detection/     # Object detection scripts
│   ├── image.py         # Process static images
│   ├── webcam.py        # Webcam stream detection
│   └── phone.py         # Phone camera stream detection
├── pose-estimation/     # Pose estimation scripts
│   └── webcam_pose.py   # Real-time pose estimation
├── data/                # Input data directory
├── output/              # Output directory for results
├── *.pt                 # Pre-trained YOLO model weights
└── requirements.txt     # Python dependencies

Models Included

The repository includes several pre-trained YOLO model weights:

Object Detection Models:

• yolov8n.pt - YOLOv8 Nano
• yolov8m-oiv7 - YOLOv8 Medium trained on Open Images V7
• yolo12n.pt - YOLO12 Nano

Pose Estimation Models:

• yolov8n-pose.pt - YOLOv8 Nano Pose
• yolo11n-pose.pt - YOLO11 Nano Pose

Instance Segmentation Models:

• yolov8n-seg.pt - YOLOv8 Nano Segmentation
• yolo11n-seg.pt - YOLO11 Nano Segmentation

Installation

1. Clone the repository:

git clone <repository-url>
cd yolo-experiment

2. Install dependencies:

pip install -r requirements.txt

Requirements

• ultralytics
• opencv-python
• pillow
• streamlit (required for the web application)

Usage

Object Detection

1. Image Detection

Process a single image with YOLO object detection:

python object-detection/image.py

• Place your input image at data/image.jpg
• Output will be saved to data/output.jpg
• Displays detection results in a window

2. Webcam Detection

Real-time object detection using your computer's webcam:

python object-detection/webcam.py

Features:

• Real-time object detection with threading for optimal performance
• FPS counter overlay
• Automatic video recording to output/webcam/recording_<timestamp>.mp4
• Press 'q' to quit

3. Phone Camera Detection

Stream and detect objects from your phone camera:

python object-detection/phone.py

Features:

• Connects to phone camera via IP webcam app (default: https://192.168.20.19:8080/video)
• Multi-threaded frame capture for reduced latency
• Frame rotation support (configurable via rot variable)
• Saves recording to output/phone/recording_<timestamp>.mp4
• Press 'q' to quit

Setup:

1. Install an IP webcam app on your phone (e.g., "IP Webcam" for Android)
2. Update the phone_url variable in phone.py with your phone's IP address
3. Ensure phone and computer are on the same network

Pose Estimation

Webcam Pose Estimation

Real-time human pose estimation using your webcam:

python pose-estimation/webcam_pose.py

Features:

• Detects 17 body keypoints (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles)
• Real-time FPS display
• Press 'q' to quit

Streamlit Web Application

Interactive web-based interface for running YOLO models with real-time webcam processing:

streamlit run app.py

Features:

• Web-based UI accessible via browser
• Three application modes:
  • Object Detection - Detect and classify objects in real-time
  • Pose Recognition - Track human body keypoints
  • Instance Segmentation - Segment and identify individual object instances
• Configurable settings via sidebar:
  • Model selection (mode-specific models available)
  • Confidence threshold slider (0.0-1.0)
  • IoU (Intersection over Union) threshold slider (0.0-1.0)
• Real-time webcam stream processing
• Live preview of processed frames

Available Models:

• Object Detection: yolo12n.pt, yolov8n.pt, yolov8m-oiv7
• Pose Recognition: yolo11n-pose.pt, yolov8n-pose.pt
• Instance Segmentation: yolo11n-seg.pt, yolov8n-seg.pt

Requirements: In addition to the base requirements, you'll need:

pip install streamlit

The app automatically opens in your default browser at http://localhost:8501

Configuration

Model Selection

Each script uses a specific YOLO model. To change the model, edit the model initialization line:

model = YOLO("yolov8n.pt")  # Change to desired model

Detection Confidence

Adjust the confidence threshold in the detection scripts:

results = model(frame, conf=0.5)  # Change conf value (0.0-1.0)

Class Filtering

To detect specific object classes only (see webcam.py example):

results = model(frame, conf=0.5, classes=[41, 45, 64, 66])  # Cup, bowl, keyboard, mouse

Performance Optimization

The webcam and phone camera scripts use threading to separate frame capture from inference, reducing latency and improving FPS. Key optimizations:

• Frame queue with max size 2 to prevent buffering lag
• Separate capture thread for continuous frame grabbing
• Minimal buffer size in OpenCV capture

License

See LICENSE file for details.