Semantic Segmentation with Transfer Learning for Off-Road Autonomous Driving.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep learning requires large datasets, often unavailable for specialized domains like off-road driving.
• Transfer learning is a trending technique but faces challenges like network size and domain differences.

Purpose of the Study:

• To investigate effective transfer learning strategies for semantic segmentation in off-road environments.
• To address limitations of large pre-trained models and explore the utility of synthetic data.

Main Methods:

• Developed a lightweight network, half the size of the original DeconvNet, for semantic segmentation.
• Employed transfer learning by fine-tuning the lightweight network with knowledge from a pre-trained DeconvNet.
• Utilized synthetic datasets as an intermediate domain before training on real-world off-road data (Freiburg Forest dataset).

Main Results:

• The proposed lightweight network achieved effective knowledge transfer.
• Fine-tuning with a realistic synthetic dataset significantly improved segmentation accuracy for real-world off-road data compared to direct transfer learning.
• Identified that overly simple or random synthetic datasets can lead to negative transfer, hindering performance.

Conclusions:

• A lightweight network combined with carefully curated synthetic data offers a viable approach for transfer learning in off-road semantic segmentation.
• Realistic synthetic data generation is crucial for successful transfer learning, while simplistic data can be detrimental.