Bridging the Domain Gap: Enhancing Underwater Laser Stripe Segmentation with Synthetic Data
Keywords: synthetic data, across-domain generalization, underwater mapping, laser segmentation
TL;DR: We introduce a synthetic dataset for underwater laser strip segmentation, demostrating the effectiveness of synthetic data to enhance model generalization for unseed marine environments
Abstract: Underwater 3D mapping using low-cost scanning systems relies on accurate laser stripes segmentation. However, the scarcity of annotated data and the variability of underwater environments limit the model's generalization and scalability. To address this issue, we introduce a synthetic dataset specifically designed for laser stripe segmentation. Created using a custom laser scanner module integrated into Blender and the Infinigen procedural generator. The dataset contains over 1,200 high-resolution images across 23 diverse terrains, each with ground truth. We evaluate the impact of synthetic data using a segmentation network trained under different field-to-synthetic data ratios. Our results show that augmenting field datasets with synthetic images significantly improves performance on unseen domains—achieving up to 10% higher recall and 7% higher precision on deep-sea imagery from the Salas y Gómez Ridge, a location with different lighting, seafloor composition, and depth. Our findings highlights the value of synthetic data for domain diversity, reducing annotation costs and enhancing model generalization, supporting broader and more robust deployment of underwater laser mapping systems.
Submission Number: 64
Loading