Machine Learning Course Report
Machine Learning Final Report: Semantic Segmentation for Autonomous Driving
Topic: Semantic segmentation as a core perception module for autonomous driving, with attention to robustness and deployment constraints.
Summary
Semantic segmentation provides pixel-level understanding of driving scenes (roads, lanes, vehicles, pedestrians, signs), making it a key component for safety-critical planning and decision-making. This report reviews representative model families and practical challenges, and summarizes directions that improve robustness, efficiency, and reliability in real-world conditions.
What’s inside
- Motivation: why pixel-level perception matters for safety and planning
- Model families and design patterns: encoder–decoder pipelines and representative segmentation approaches
- Practical challenges:
- weather / illumination shift and domain generalization
- real-time constraints and efficiency
- small-object segmentation
- annotation scarcity and label noise
- class imbalance
- multimodal fusion (optional)
- interpretability and reliability
Report