SegLight - Light Semantic Segmentation
Project Overview
SegLight is a lightweight semantic segmentation model designed specifically for humanoid soccer robots. The project focuses on implementing a high-performance semantic segmentation solution with minimal inference time on CPU, making it ideal for real-time robotic applications where computational resources are limited.
The network architecture was inspired by previous team projects and optimized for deployment on embedded systems used in humanoid soccer robots. The model achieves excellent performance while maintaining low computational overhead, enabling real-time scene understanding for autonomous navigation and decision-making.
Key Features
- Lightweight architecture optimized for CPU inference
- Real-time semantic segmentation for robotic applications
- Custom dataset preparation and augmentation pipeline
- Support for TensorFlow and Keras model formats
- C++ inference support via Cppflow
- Flexible color-map configuration for different semantic classes
- Data augmentation with repeat functionality for small datasets
Technology Stack
TensorFlow & Keras
Deep learning framework for building and training the semantic segmentation model with version compatibility for both Keras v2 and v3
Python
Primary programming language for model training, data preprocessing, and augmentation pipeline
C++ & Cppflow
Used for efficient model inference on embedded systems and real-time robotic applications
Results & Impact
The SegLight model successfully achieves real-time semantic segmentation performance on CPU-based systems, making it highly suitable for deployment on humanoid soccer robots. The lightweight architecture ensures minimal computational overhead while maintaining accuracy for critical scene understanding tasks. The project has been successfully integrated into robotic systems and demonstrates practical applicability in resource-constrained environments.
Technical Highlights
- Support for multiple model formats (.h5, .keras, .pb)
- Optimized for TensorFlow 2.16+ with legacy Keras support
- Flexible dataset structure with series-based organization
- Custom augmentation pipeline for handling small datasets
- Compatible with humanoid soccer robot platforms
Project Gallery