This project leverages the Segformer pretrained model for billboard segmentation via semantic segmentation, with a focus on Indian billboards. By utilizing transfer learning on a custom dataset, the model precisely classifies and outlines billboards within images, facilitating efficient detection and analysis. The methodology incorporates several key improvements, including:

• Dataset Augmentation to increase model robustness.
• Data Balancing to address class imbalance issues.
• Advanced Post-Processing techniques to refine segmentation results.

These techniques significantly enhance segmentation accuracy, addressing challenges commonly encountered in binary semantic segmentation tasks.

The dataset is structured as follows:

├── Data
│   ├── Train
│   │   ├── images  # Training images
│   │   ├── masks   # Corresponding binary masks
│   ├── Validation
│   │   ├── images  # Validation images
│   │   ├── masks   # Corresponding binary masks

You can download the dataset from the following link:

• Billboard Dataset

To train the Segformer model for billboard segmentation, you can follow along with the Google Colab notebook linked below. The notebook provides step-by-step instructions for data preprocessing, model training, and evaluation.

The following graph shows the training and validation performance over time, demonstrating the model's convergence and improvement in segmentation accuracy:

Below are sample outputs of the model on test images. These showcase the model's ability to accurately segment billboards:

Here are some final outputs after post-processing, showing the refined billboard segmentation:

This README is focused solely on the training aspect of the project. A separate README file will cover API-related details and deployment.