First Look: YOLOv5

In my technical exploration, I recently came across YOLOv5, a powerful tool in the world of computer vision. This discovery stemmed from an experiment on Replit, where I aimed to develop an image recognition application for users to upload photos of their personal libraries. The concept was straightforward: an image of bookshelves would be processed via OCR (Optical Character Recognition) or similar text-identification techniques. The identified books would then be cross-referenced with book marketplaces to generate a CSV file estimating the library’s value.

"Let me begin with whatever subject I please, for all subjects are linked with one another." — Michel de Montaigne

During this process, one significant challenge arose: recognizing text on vertically aligned book spines. The limitations of the OCR approach led me to explore object detection frameworks, eventually landing on YOLOv5 (You Only Look Once, version 5).

What is YOLOv5?

YOLOv5 is an open-source, deep learning model designed for real-time object detection. Built for speed and accuracy, YOLOv5 excels at identifying, classifying, and localizing objects within images and video streams. Its PyTorch-based implementation is beginner-friendly and supports both pre-trained models and custom training for domain-specific applications.

Initial Testing with YOLOv5

My first tests leveraged YOLOv5’s detect.py script on a dataset of four images, with the detection confidence threshold set to 40%. The results were intriguing. YOLOv5 performed exceptionally well in detecting people but struggled with non-standard objects like vertically aligned book spines. This was expected, as I was using the lightweight yolov5s model, which prioritizes speed over accuracy.

YOLOv5 Model Variants:

• yolov5s: Small and fast, but less accurate.
• yolov5m: Medium-sized, balancing speed and accuracy.
• yolov5l: Larger, more accurate, but slower.
• yolov5x: The largest and most accurate, optimized for precision over speed.

I haven’t yet explored training the model on a custom dataset or experimented with the larger, more accurate variants. This first look focused solely on the pre-trained yolov5s model.

Observations and Challenges

While YOLOv5's person detection impressed me, there were some humorous misclassifications. For instance:

1. The centerpiece of Michelangelo’s iconic ceiling mural was identified as a "person" and, inexplicably, a "sandwich." Take a look here.
2. A scene from Jurassic Park's "Journey to the Island" was unable to classify the helicopter—highlighting the absence of a COCO dataset category for helicopters.

Despite these object detection errors, YOLOv5’s speed and flexibility stand out. Running the model on an older, compatible Python version (Python 3.10.16) in a virtual environment configured specifically for PyTorch showcased its lightweight implementation and ease of deployment.

Conclusion: First Impressions

This initial foray into YOLOv5 demonstrates its strengths as a robust and fast object detection model. While it requires customization and training for niche use cases, its out-of-the-box functionality is impressive, particularly for detecting people. My next steps involve fine-tuning the model for library-specific object detection, experimenting with larger variants like yolov5l, and exploring its real-time capabilities in video feeds.

For now, YOLOv5 proves to be a fascinating and highly capable tool for any developer venturing into object detection, offering an accessible entry point into the cutting-edge world of computer vision.