Comparative Analysis of Custom YOLOv8 Backbones

optimizing YOLOv8 backbones for parking space detection

This project focuses on optimizing YOLOv8 backbone architectures for parking space detection using the PKLot dataset. We explored multiple lightweight and high-performance backbone models to achieve better precision–recall trade-offs, reduced inference latency, and enhanced computational efficiency for real-time applications.

Project Page

You can find the project pagehere.

Highlights

  • Custom Backbone Architectures: ResNet-18, VGG16, EfficientNet-B0, and GhostNet-P2 modifications.
  • Performance Metrics: Achieved a mAP@0.5:0.95 of 98.6%, outperforming baseline YOLOv8.
  • Deployment Ready: Optimized for real-time edge devices with reduced model size and faster inference.
  • Dataset: PKLot dataset of parking lot images with high variability in lighting, angle, and occlusion.

Key Contributions

  • Designed and implemented custom lightweight backbones tailored for constrained hardware.
  • Conducted extensive ablation studies analyzing accuracy, latency, and resource usage.
  • Developed a streamlined training pipeline for rapid prototyping and benchmarking.

References