<![CDATA[Maritime obstacle detection is a critical challenge for Unmanned Surface Vehicles (USVs) operating in complex and dynamic environments. This study investigates the effectiveness of fine-tuning Panoptic FPN, a Mask R-CNN-based architecture augmented with Feature Pyramid Networks, for panoptic segmentation on the LaRS (Lake, River, Seas) dataset. Unlike prior work that explored model comparisons broadly, this research focuses specifically on the impact of hyperparameter tuning and backbone selection on maritime panoptic segmentation performance. Through systematic ablation studies, we demonstrate that adjusting the learning rate to 0.002 and the gamma decay factor to 0.2 yields significant improvements. Our fine-tuned Panoptic FPN with a ResNet-50 backbone achieves a Panoptic Quality (PQ) of 45.31%, surpassing the previous state-of-the-art Mask2Former Swin-B (41.7%) by 3.61 percentage points. Notably, ResNet-50 outperforms the deeper ResNet-101 backbone (36.47% PQ), suggesting that heavier architectures may overfit on domain-specific maritime datasets. Furthermore, Panoptic FPN requires only 8 hours of training compared to approximately 2 days for Mask2Former Swin-L, demonstrating superior computational efficiency. These findings highlight that targeted fine-tuning of lightweight architectures can outperform larger transformer-based models in maritime panoptic segmentation tasks.]]>
Copyrights © 2026
