🧠 Daily Study Log [2025-06-19]

Today was all about combining hands-on computer vision tasks with core CNN theory.
I trained a segmentation model on real urban scenes and revisited a sign language classification task to solidify my understanding of CNN fundamentals.

🏙️ Image Segmentation — Urban Scene Understanding

• Objective: Assign semantic labels (road, sky, building, etc.) to every pixel in an image — essential for self-driving perception.
• Dataset: Cityscapes (_leftImg8bit.png, _gtFine_labelIds.png)
• Model: Unet(resnet34) from segmentation_models_pytorch
• Input Size: Resized to 256x256
• Loss Function: CrossEntropyLoss() with classes=34
• Training: 5 epochs — loss steadily decreased
• Result: Predicted masks matched ground truth very well; roads, skies, cars, and buildings were segmented accurately

✅ Takeaways

• I made a mistake in setting the number of output classes, which helped me understand how CrossEntropyLoss() behaves — a valuable learning moment.
• Semantic segmentation gives visual feedback, which made model performance much more intuitive to grasp.

🧪 CNN Basics Recap — Sign Language Classifier

• Task: Classify images of hand gestures into 10 sign language categories
• Model: Simple CNN (Conv → ReLU → MaxPool → Dense)
• Input: RGB images resized to 224x224
• Loss: Categorical CrossEntropy
• Augmentation: Horizontal flips, random rotations using ImageDataGenerator
• Performance Check: Printed classification report and plotted training accuracy

✅ Reflections

• The simple CNN worked decently and was great for reviewing layer flow and preprocessing steps.
• Reinforced the importance of proper image labeling and class balancing.

🎯 Next Steps

• Try advanced architectures like SegFormer and DeepLab v3+
• Evaluate using mIoU, pixel accuracy, and visualization tools
• Build a proper inference pipeline → Prepare for deployment

✅ TL;DR

📍 U-Net: Deployed on urban segmentation
📍 CNN: Refreshed using sign language dataset
📍 Next: Test SegFormer + Evaluate with mIoU