OilGaze: A Ground-View Dataset for Robotic Oil-Spill Inspection

Oil-spill detection research has largely focused on satellite and aerial imagery, leaving a gap for close-range shoreline inspection from the perspective of a human or ground robot. We introduce ShoreOil-GV, a ground-view RGB dataset for visible surface oil segmentation on Peruvian beaches affected by the La Pampilla oil spill. The dataset captures real sandy and rocky shoreline environments, including challenging visual distractors such as wet sand, shadows, and dark rocks. Data are organized by acquisition sequence and site to enable evaluation across different locations and environmental conditions while reducing leakage from adjacent video frames. Using ShoreOil-GV, we establish a semantic segmentation benchmark comparing fine-tuned U-Net, DeepLabV3+, and SegFormer models with zero-shot SAM 3. Our results show that, although state-of-the-art foundation models provide strong general-purpose visual representations, zero-shot segmentation remains insufficient for reliable oil delineation in this challenging shoreline domain. ShoreOil-GV provides a benchmark for evaluating domain-specific adaptation and advancing perception systems for autonomous environmental monitoring.