Real Time Semantic Segmentation using Efficient Neural Network
Project Overview
Semantic segmentation is a computer vision task where every pixel in an image is classified into a specific category. This project implements ENet, a lightweight deep neural network optimized for real-time segmentation with low latency. The model is tested on the CamVid dataset, and has achieved efficient segmentation performance at 10 FPS with a pixel accuracy of 75%.
Languages & Tech stack
- Python for implementation.
- PyTorch for Deep Learning framework.
- CamVid for dataset.
Methodology
- Model Architecture
- ENet uses an encoder-decoder structure, optimized for speed and efficiency.
- Early downsampling and dilated convolutions help reduce the number of computations.
- Implementation
- The model is trained using 367 training images, validated on 101 images, and tested on 233 images.
- Hyperparameter tuning: Learning rate (5×10⁻⁴), batch size (5), weight decay (2×10⁻⁴).
- Training & Testing
- The model is trained for 100 epochs, and the loss converged effectively.
- The segmentation masks were evaluated using Pixel Accuracy, IoU, and F1 Score.
Results
- Per-pixel accuracy - 75%.
- IoU - 0.51.
- F1 score - 0.64.
Project Snippets
-
Segmentation Result
