Skip to content

YOLO_SELECTION_GUIDE.md

Latest commit

 

History

History
201 lines (147 loc) · 5.37 KB

YOLO_SELECTION_GUIDE.md

File metadata and controls

201 lines (147 loc) · 5.37 KB

YOLO Selection Guide: V8 vs V3

This guide helps you choose between YOLO V8 (Ultralytics) and YOLO V3 (Darknet) for your 3D detection system.

Quick Comparison

Feature YOLO V8 (Ultralytics) YOLO V3 (Darknet)
Accuracy ⭐⭐⭐⭐⭐ (10-20% higher mAP) ⭐⭐⭐⭐
Speed ⭐⭐⭐⭐⭐ (2-3x faster) ⭐⭐⭐
Model Size ⭐⭐⭐⭐⭐ (30x smaller) ⭐⭐
GPU Support ✅ PyTorch/CUDA ✅ CUDA
Active Development ✅ Yes ⚠️ Legacy
Ease of Use ✅ Easy ⚠️ Requires weights download

When to Use YOLO V8 (Recommended)

✅ Use YOLO V8 if:

  • You want better accuracy and speed
  • You have limited storage (smaller models)
  • You want the latest features and updates
  • You're starting a new project
  • You need faster inference

When to Use YOLO V3 (Legacy)

⚠️ Use YOLO V3 if:

  • You need compatibility with existing systems
  • You have pre-trained custom weights for YOLO V3
  • You're maintaining legacy code
  • You have specific requirements for Darknet framework

Usage

Option 1: YOLO V8 (Default)

# Default (YOLO V8 Medium)
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py

# Or explicitly specify YOLO V8
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=true

# Choose YOLO V8 model variant
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=true yolo_model:=yolov8n
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=true yolo_model:=yolov8m
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=true yolo_model:=yolov8l

Option 2: YOLO V3 (Legacy)

# Use YOLO V3
ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=false

# Or use dedicated YOLO V3 launch file
ros2 launch darknet_ros_3d yolov3_3d.launch.py

Option 3: Direct Launch Files

# YOLO V8 only
ros2 launch darknet_ros_3d yolov8_3d.launch.py

# YOLO V3 only
ros2 launch darknet_ros_3d yolov3_3d.launch.py

Model Variants (YOLO V8 Only)

Model Size mAP Speed Use Case
yolov8n 6.2MB 37.3 80+ FPS Speed priority
yolov8s 22MB 44.9 60+ FPS Balanced
yolov8m 52MB 50.2 40+ FPS Recommended
yolov8l 87MB 52.9 25+ FPS Accuracy priority
yolov8x 136MB 53.9 20+ FPS Best accuracy
yolov11m 40MB 51.5 45+ FPS Latest generation

Performance Comparison

YOLO V8 Medium vs YOLO V3

  • Accuracy: YOLO V8m has ~10-20% higher mAP
  • Speed: YOLO V8m is 2-3x faster on same hardware
  • Model Size: YOLO V8m (52MB) vs YOLO V3 (235MB)
  • Memory: YOLO V8m uses less GPU memory

Configuration

YOLO V8 Configuration

Edit: src/ultralytics_ros/config/yolov8m.yaml

yolo_detector_node:
  ros__parameters:
    model_path: 'yolov8m.pt'
    device: 'cuda:0'  # or 'cpu'
    confidence_threshold: 0.25
    input_topic: '/camera/camera/color/image_raw'

YOLO V3 Configuration

Edit: src/darknet_ros/darknet_ros/config/yolov3.yaml

darknet_ros:
  ros__parameters:
    yolo_model:
      config_file:
        name: yolov3.cfg
      weight_file:
        name: yolov3.weights
      threshold:
        value: 0.3

Switching Between Versions

From YOLO V3 to YOLO V8

  1. Install dependencies:

    pip3 install -r src/ultralytics_ros/requirements.txt

  2. Build package:

    colcon build --packages-select ultralytics_ros

  3. Launch with YOLO V8:

    ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=true

From YOLO V8 to YOLO V3

  1. Ensure YOLO V3 weights are downloaded:

    # Check if weights exist
ls src/darknet_ros/darknet_ros/yolo_network_config/weights/yolov3.weights

# If not, download:
cd src/darknet_ros/darknet_ros/yolo_network_config/weights/
wget https://pjreddie.com/media/files/yolov3.weights

  2. Launch with YOLO V3:

    ros2 launch darknet_ros_3d darknet_ros_3d_complete.launch.py use_yolov8:=false

Troubleshooting

YOLO V8 Issues

Problem: Model not loading

  • Solution: Models auto-download on first run. Check internet connection.

Problem: CUDA out of memory

  • Solution: Use smaller model (yolov8n) or reduce image_size in config

YOLO V3 Issues

Problem: Weights file not found

Problem: CUDA compilation errors

  • Solution: See ENABLE_GPU.md for CUDA setup instructions

Recommendations

For New Projects

👉 Use YOLO V8 - Better performance, easier setup, active development

For Existing Projects

  • If already using YOLO V3: Can continue using it
  • If migrating: Consider YOLO V8 for better performance

For Production

  • YOLO V8 Medium (yolov8m) - Best balance of accuracy and speed
  • YOLO V8 Large (yolov8l) - If accuracy is critical

For Development/Testing

  • YOLO V8 Nano (yolov8n) - Fastest, good for quick testing

Summary

Default Choice: YOLO V8 (Ultralytics)

  • Better accuracy and speed
  • Smaller models
  • Easier to use
  • Active development

Legacy Choice: YOLO V3 (Darknet)

  • Use only if you have specific requirements
  • Or for compatibility with existing systems