Clothing Color Classifier 👕🎨

Introduction

Can a small edge AI model learn to recognize colors on clothing items just from a webcam feed — with no cloud, no heavy hardware, and a small custom dataset? Clothing Color Classifier answers that question with a resounding yes.

This project uses Edge Impulse Studio to train an Object Detection model that locates a clothing item's colored area (in this case, the chest/bib section of small kitty toy used as color samples) and classifies its color into one of four categories: Red, Magenta, Pink, and Sky Blue. The final model reached a 94.7% F1 score, with real-time detection tested live through a browser webcam — no software installation required.

🛠️ Hardware & Tools

• USB Webcam connected to a laptop — the only hardware needed for data collection and live testing
• Edge Impulse Studio — for data labeling, model design, training, and browser-based live classification
• No microcontroller deployment was required for this version; everything runs directly through the browser

📸 Data Collection

Data was collected directly through Edge Impulse's Data Acquisition tool, connected to the webcam. Each sample captured a small object with a distinctly colored section, later labeled by color.

Final dataset:

• 146 images total (120 training / 26 test)
• 4 balanced classes: Magenta (30), Pink (37), Red (30), Sky Blue (37)
• Images captured with varied angles, distances, and differnet backgrounds.

🎯 Labeling Strategy

Using Object Detection labeling, each image received a tight bounding box drawn strictly around the colored region (the chest/bib area), deliberately excluding unrelated parts of the object (like the white body or decorative bow). This precision mattered a lot: boxes that were too loose picked up background color and blurred the distinction between similar shades; boxes that were too tight lost enough pixel information for the model to learn from. Finding that balance was key to final accuracy.

🧠 Impulse Design

Final training configuration:

• Training cycles: 100
• Learning rate: 0.001
• Data augmentation: enabled
• Training processor: CPU

⚡ Why FOMO?

FOMO (Faster Objects, More Objects) is Edge Impulse's ultra-lightweight object detection algorithm, built on a truncated MobileNetV2 0.35 backbone. Instead of predicting full bounding boxes like YOLO, it locates objects by predicting just their center point on a grid — using up to 30× less memory and compute, enabling real-time inference even on tiny microcontrollers. This efficiency made it the ideal choice for a small, single-object-per-image classification task like this one.

📊 Results

F1 Score: 94.7%

Getting here required iterating on both dataset balance and bounding box precision — particularly for the lighter shades, which initially struggled to stand out from light backgrounds and were confused with each other.

🔮 Live Demo

The trained model was tested directly in the browser using Edge Impulse's Live Classification feature — no downloads, no firmware flashing.

💡 Key Takeaways

• Bounding box precision has a sweet spot — too loose or too tight both hurt accuracy
• FOMO is a strong choice for small, custom Object Detection datasets compared to heavier models like YOLOv5
• A fully functional Edge AI pipeline — from data collection to real-time inference — is achievable entirely through the browser, with no dedicated hardware deployment needed

🚀 Future Work

• Deploy the trained model to an edge device for standalone, offline color detection
• Expand the color palette and test on real fabric/clothing items instead of toy samples
• Explore combining this with a "outfit matching" recommendation feature
• Explore printable paper templates: instead of relying only on physical plush toys, a simple black-and-white template of the object could be printed multiple times and hand-colored in different shades, offering a fast, low-cost way to expand color variety in the dataset without needing more physical items

🔗 Try the live model :

https://studio.edgeimpulse.com/public/1116206/live