Comparative Benchmarking of Object Detection Models: YOLOv8 Nano, SSD MobileNet, and Faster R-CNN

📖 Overview

This project implements and evaluates three state-of-the-art object detection models — YOLOv8 Nano, SSD MobileNet, and Faster R-CNN — on the COCO128 dataset.

The goal is to compare performance, speed, and accuracy trade-offs between lightweight, mobile-friendly, and high-precision architectures. The results highlight which models are best suited for real-time detection, mobile deployment, and precision-focused offline tasks.

---

🎯 Objectives

• Select a real-world dataset (COCO128) relevant for everyday object detection.
• Apply data preprocessing (resizing, normalization, train-validation split).
• Implement three different architectures: YOLOv8, SSD MobileNet, and Faster R-CNN.
• Evaluate performance using precision, recall, mAP50, mAP50-95, and inference time.
• Conduct a comparative analysis with case-specific recommendations.

---

📂 Dataset

• Dataset Used: COCO128
• Lightweight subset of the COCO dataset (128 images, 929 annotations).
• Covers diverse object categories: people, vehicles, animals, and everyday items.

---

⚙️ Models Evaluated

1. YOLOv8 Nano

• Type: Lightweight, speed-optimised version of YOLOv8
• Strengths: High confidence for distinct objects (e.g., Cat 0.94, Couch 0.91), efficient computational cost, suitable for real-time GPU deployments
• Weaknesses: Moderate recall, lower confidence on small/overlapping objects (e.g., remotes <0.34)
• Inference Time (CPU): ~0.64s/image

2. SSD MobileNet

• Type: Mobile-optimised Single Shot Detector
• Strengths: Very fast on CPU/GPU, lightweight for mobile or edge deployments
• Weaknesses: Struggles with accuracy on complex or small objects (e.g., Cat detected as Egyptian cat, 0.29 confidence)
• Inference Time (CPU): ~3.8s/image

3. Faster R-CNN

• Type: Region-based Convolutional Neural Network
• Strengths: High accuracy with precise bounding boxes
• Weaknesses: Slow inference (~9.5s/image), occasional misclassifications (remote → microwave at 0.99–1.00 confidence)
• Best Use Case: Offline or batch processing when accuracy is critical

---

📊 Results Summary

Model	Inference Speed (CPU)	Accuracy	Key Observations
YOLOv8 Nano	~0.64s/image	High	Balanced speed & accuracy; strong on distinct objects, weaker on small/overlapping ones
SSD MobileNet	~3.8s/image	Moderate	Very lightweight; trades precision for speed
Faster R-CNN	~9.5s/image	High	Most accurate but slow; misclassifications noted

Run Notebook in Google Colab