Saliency-Guided Domain Adaptation for Left-Hand Driving in Autonomous Steering

TL;DR

This work tackles the domain shift encountered when deploying right-hand driving autonomous steering models in left-hand driving environments, using Australian highways to evaluate adaptation. It proposes a flipped U.S. data pretraining strategy followed by fine-tuning on local data and validates this approach on PilotNet and ResNet, complemented by saliency-map analysis to track attention shifts. The key finding is that flipping before fine-tuning yields the best performance, with substantial improvements in steering accuracy and a pronounced shift of attention toward left-side road cues, and this trend holds across architectures. The method offers a simple, data-efficient pathway to improve cross-regional deployment of end-to-end AV steering with minimal retraining requirements.

Abstract

Domain adaptation is required for automated driving models to generalize well across diverse road conditions. This paper explores a training method for domain adaptation to adapt PilotNet, an end-to-end deep learning-based model, for left-hand driving conditions using real-world Australian highway data. Four training methods were evaluated: (1) a baseline model trained on U.S. right-hand driving data, (2) a model trained on flipped U.S. data, (3) a model pretrained on U.S. data and then fine-tuned on Australian highways, and (4) a model pretrained on flipped U.S. data and then finetuned on Australian highways. This setup examines whether incorporating flipped data enhances the model adaptation by providing an initial left-hand driving alignment. The paper compares model performance regarding steering prediction accuracy and attention, using saliency-based analysis to measure attention shifts across significant road regions. Results show that pretraining on flipped data alone worsens prediction stability due to misaligned feature representations, but significantly improves adaptation when followed by fine-tuning, leading to lower prediction error and stronger focus on left-side cues. To validate this approach across different architectures, the same experiments were done on ResNet, which confirmed similar adaptation trends. These findings emphasize the importance of preprocessing techniques, such as flipped-data pretraining, followed by fine-tuning to improve model adaptation with minimal retraining requirements.

Figures (8)

• Figure 1: Training Methodology: Pretraining on original/flipped U.S. data (top) and subsequent fine-tuning on Australian highway data (bottom)
• Figure 2: Image preprocessing steps for a random image, including cropping the top 400 pixels to remove irrelevant areas, resizing to 200 × 66 for consistency with the NVIDIA dataset, and converting to YUV color space to improve road feature extraction.
• Figure 3: Region masks (Left, Center, Right) generated based on lane boundary detection for a random image to analyze saliency spread across different areas of the road without applying ROI
• Figure 4: A random image with ROI highlighted in red and corresponding segmented road regions(Left, Center, Right) after ROI application
• Figure 5: MSE Comparison Across Four Training Strategies - The highest MSE (930.03) from trained on flipped U.S. data versus the lowest MSE (3.51) achieved by combining flipped data pretraining with fine-tuning demonstrates that successful domain adaptation requires both appropriate pretraining and fine-tuning for left-hand driving conditions.