🏙️ [2025-06-19] Urban Scene Segmentation & CNN Recap

Today was a productive day filled with hands-on practice in image segmentation and a solid CNN refresher.
I trained a U-Net model to predict pixel-wise masks on real urban scenes and verified that the predictions matched the ground truth closely.
In parallel, I revisited the classic Dog vs. Cat classifier to reinforce CNN basics.

🧱 1. Image Segmentation — Urban Street Scenes

This project tackles semantic segmentation on urban environments using the Cityscapes dataset.
The goal was to assign each pixel a semantic label (e.g., road, sky, building) — a key component for autonomous driving systems.

✅ Technical Highlights

• Dataset: Cityscapes (_leftImg8bit.png, _gtFine_labelIds.png)
• Model: Unet(resnet34) from segmentation_models_pytorch
• Loss Function: CrossEntropyLoss() with classes=34
• Input Size: Resized to 256x256
• Training: 5 epochs with steadily decreasing loss
• Evaluation: Visual comparison between predicted and ground truth masks showed strong alignment

🖼️ Sample Visualization

• Displayed side-by-side plots of Input Image, Ground Truth Mask, and Predicted Mask
• Road, buildings, sky, and vehicles were segmented accurately

🧪 2. CNN Recap — Dog vs. Cat Classifier

As a complementary activity, I reviewed a basic CNN model using the Dog vs. Cat dataset.
This helped reinforce my understanding of convolutional layers, activation functions, and data preprocessing.

✅ Key Elements

• Input: RGB images (224x224)
• Model: Basic CNN — Conv → ReLU → MaxPool → Dense
• Loss: Binary CrossEntropy
• Augmentation: Horizontal flip, rotation using ImageDataGenerator
• Performance Check: Visualized accuracy and loss curves

💡 Reflections

• Segmentation offers much more visual intuition than plain classification.
• I made a mistake in setting the number of output classes, which helped me fully understand how CrossEntropyLoss() works — lesson learned.
• It was satisfying to see how well the predicted mask aligned with the ground truth.
• CNN fundamentals are solid now, and I can clearly see the value of transfer learning for more efficient workflows.

🎯 Final Thoughts

This session combined practical low-level vision tasks with high-level architectural understanding.
Next, I plan to experiment with SegFormer, DeepLab v3+, and evaluate models using mIoU, pixel accuracy, and visual comparisons.

✅ Segmentation: Deployed and working
🐶 CNN: Refreshed and reinforced
📈 Next Step: Validation loop, mIoU scoring, model deployment