This repository contains an implementation of a semantic segmentation pipeline using the DeepLabV3+ architecture. The project focuses on leveraging deep learning to perform pixel-level classification, enabling precise object boundaries detection. The model is built using the segmentation-models-pytorch library, showcasing an efficient workflow for training and evaluating deep computer vision models.
- Architecture: DeepLabV3+ – a state-of-the-art model for semantic segmentation.
- Backbone: ResNet-101 encoder for robust feature extraction.
- Transfer Learning: Pre-trained ImageNet weights were used to accelerate convergence and improve accuracy.
- Hardware Acceleration: Optimized for NVIDIA T4 GPU (trained in Google Colab).
- Modern Formats: Implementation utilizes safetensors for secure and fast model weight loading.
The following image demonstrates the model's ability to identify and segment various object classes within a scene.
The training loss curve indicates a steady convergence, confirming that the model successfully learned the features from the dataset.
