📸 [2025-06-08] Real-Time Activity Recognizer — Final Wrap-up

After weeks of crawling, preprocessing, modeling, and fine-tuning, this marks the final wrap-up of the Real-Time Daily Activity Recognizer project.
From simple grayscale CNNs to EfficientNetB0 with dropout and augmentation — this journey was a hands-on dive into the full image classification pipeline.

📚 1. What I Built

This project combined image crawling, model training, and webcam integration to simulate a real-world computer vision task.
I handled everything from data collection to inference — using Keras, OpenCV, and transfer learning.

🧱 Technical Highlights

• Google Image crawler using selenium & urllib
• Image classification model using EfficientNetB0 + Dropout(0.5)
• Real-time prediction via webcam (cv2.VideoCapture)
• Strong data augmentation using ImageDataGenerator
• Full training loop with callbacks: checkpointing, early stopping, LR scheduler
• Visualization with matplotlib (loss/accuracy curves, saved to /figures)

💡 Reflections

• Most important lesson: Data quality matters more than model architecture.
  → Crawling images manually had too many noisy, irrelevant samples — next time I’ll use ImageNet, Kaggle datasets, or structured action datasets.

• Model-wise, EfficientNetB0 + dropout gave the best trade-off between training speed and generalization.

• Technically, integrating OpenCV inference was super fun — it gave an actual feeling of “real-time AI”.

• I’m proud to have taken this from concept → dataset → model → live prediction.

🎯 Final Thoughts

This was never meant to be production-grade — but it taught me the entire pipeline.
From crawling to inference, from augmentation to callbacks, I got to see where things break and how to improve them.

Now it’s time to move on to new projects, with better data and deeper models.

✅ Project: Completed
🧠 Lessons: Internalized
🔥 Next stop: Something bigger.