Object Detection in 20 Years: A Survey

Analysis

Why This Paper Matters

Object detection is a cornerstone of computer vision, enabling applications from autonomous driving to medical imaging. This survey, published in Proceedings of the IEEE and cited over 3,700 times, provides a definitive historical and technical overview of the field's 25-year evolution. Its significance lies in bridging the gap between early hand-crafted approaches and modern deep learning-based detectors, offering a coherent narrative that helps researchers understand how and why the field has progressed.

By covering milestones, datasets, metrics, and architectural building blocks, the paper serves as both an educational resource and a reference for practitioners. It contextualizes the deep learning revolution within the broader history of computer vision, highlighting the ingenious thinking of early pioneers and the long-term perspective that shaped today's techniques. This historical framing is particularly valuable for newcomers and for those seeking to identify future research directions.

Technical Contributions

• Comprehensive historical review: Traces object detection from the 1990s to 2022, identifying key milestones such as Viola-Jones, HOG+SVM, DPM, and the R-CNN family, YOLO, SSD, and transformer-based detectors.
• Taxonomy of detection systems: Organizes fundamental building blocks including backbone networks, neck designs, and detection heads, providing a structured understanding of modern architectures.
• Dataset and metric survey: Reviews major benchmarks (e.g., PASCAL VOC, COCO, Open Images) and evaluation metrics (e.g., mAP, IoU), discussing their evolution and impact.
• Speedup techniques: Covers methods for improving detection efficiency, such as lightweight backbones, pruning, quantization, and specialized kernels.
• State-of-the-art overview: Summarizes recent advances, including one-stage and two-stage detectors, anchor-free methods, and transformer-based approaches.

Results

As a survey, the paper does not introduce new experimental results. Instead, it consolidates performance trends from the literature, showing the dramatic improvements in accuracy and speed achieved by deep learning detectors compared to traditional methods. For example, it discusses the progression from R-CNN's mAP on PASCAL VOC to modern detectors achieving over 50% mAP on COCO, and the reduction in inference time from seconds to milliseconds. The paper also highlights the impact of larger datasets and improved evaluation protocols on driving progress.

Significance

The survey's significance extends beyond its comprehensive coverage. By providing a clear roadmap of the field's evolution, it has become a standard reference for researchers and engineers, influencing how object detection is taught and researched. Its high citation count attests to its impact on the computer vision community. Moreover, the paper's discussion of speedup techniques and recent state-of-the-art methods offers practical guidance for deploying detection systems in real-world applications. As object detection continues to evolve with transformers and self-supervised learning, this survey remains a foundational text for understanding the field's trajectory and future challenges.