A Comprehensive Review of the YOLO Series (v1 – v11)

---

📌 Overview

This repository contains my undergraduate thesis:

Topic: A Comprehensive Review of YOLO Series (v1 to v11)
Purpose: To explore the evolution of the YOLO (You Only Look Once) object detection family, review its architectural changes, and highlight the impact of those changes on real‑time object detection performance.

YOLO (You Only Look Once) is a groundbreaking series of single‑stage object detection algorithms that balance speed and accuracy.
First introduced in 2015, YOLO has evolved through multiple community‑driven iterations, each version adding new ideas while keeping real‑time detection at its core.

---

🎯 Research Goals

1. Comprehensive Review:
   Track the architectural evolution of YOLO from v1 (2015) to v11 (2025) and examine the key techniques introduced along the way.

2. Qualitative Analysis:
   Compare YOLOv6, YOLOv9, YOLOv10, and YOLOv11 on real‑time inference and accuracy using:

   • PyTorch 2 (eager execution)
   • MS COCO 2017 dataset
   • NVIDIA Tesla T4 GPU (2 vCPUs on Google Cloud)

3. Impact and Trends:
   Highlight which architectural innovations persisted, which faded, and how YOLO’s single‑stage paradigm influenced modern object detection research.

---

🔬 Key Results

• ✅ YOLOv11 achieved the best trade‑off between post‑processing time and mAP scores.
• ✅ YOLOv9 with large scaling still scored the highest mAP (55.6%), showing headroom for improvement.
• ✅ YOLOv10, with 44% fewer parameters, achieved performance close to YOLOv11 — demonstrating efficiency gains.
• ✨ Core architectural ideas such as CSPNet, backbone‑neck‑head structures, and augmentations like Mosaic and CutMix have persisted across YOLO versions.
• 🚀 YOLO remains lightweight and easily deployable across devices, making it widely adopted in real‑time systems.

---

📊 Discussion

Although YOLO continues to improve, not all techniques carried over:

• Self‑Adversarial Training (YOLOv4) and Programmable Gradient Information (PGI) didn’t persist in later versions.
• Results show that revisiting these techniques might further enhance accuracy or reduce parameters in future versions.

---

🙌 Acknowledgments

Special thanks to the YOLO authors and the open‑source computer vision community for their continuous contributions that inspired this review.