Today, I kicked off a brand-new computer vision project:
Real-Time Daily Activity Recognition based on webcam input and CNNs.

It’s my first time working on a project that’s meant to react live to human actions — like brushing teeth, eating, or typing — and I’m already hooked on the challenge.

🚀 What I Did (Start of the Project)

• 🧠 Decided on the project scope: real-time activity classification from webcam images
• 🗂️ Selected 8 daily activities: brushing_teeth, eating, drinking, typing, reading, sleeping, walking, washing_face
• 🔍 Designed precise image search keywords like "person reading book" instead of "reading" to increase quality
• 📸 Crawled 400+ images per class using Bing image search
• 🧼 Filtered corrupted or unopenable image files using PIL.Image.verify
• 📁 Saved the dataset locally under images/ directory (to later split into train/val)
• ⚠️ Took care to avoid any copyright violation by not redistributing the dataset — it’s for learning only

💭 What I Learned

• Specific keywords like "person brushing teeth" yield way better results than vague ones like "brushing"
• Real-world data is messy: you always need to clean up after crawling
• Data labeling by folder name is super convenient when working with ImageDataGenerator

🌱 What’s Next

• 🔨 Split the dataset into train/val with proper ratio (80/20)
• 🧪 Build and train a CNN using MobileNetV2 as a base
• 🎯 Evaluate initial performance and decide whether fine-tuning is needed
• 🖥️ Eventually: hook the trained model up to OpenCV for real-time webcam inference

🗓️ Reflection

It’s exciting to move beyond static classification and into something reactive.
This project feels a bit more personal — it’s recognizing the kind of daily behaviors I actually do.

Tomorrow I’ll dive into model training.
For now, the dataset is ready — and so am I.

“You can’t train a model on data you haven’t gathered.
Today was about gathering.”