Accurate monitoring of ethylene leakage is critical to safety production and pollution prevention. Addressing the current lack of publicly available datasets for infrared ethylene semantic segmentation, this paper constructs and releases a large-scale, high-quality infrared ethylene image dataset—the first specifically designed for ethylene leakage segmentation tasks. It comprises 100,000 high-resolution infrared images with pixel-level fine annotations. Based on this dataset, this paper systematically evaluates eight mainstream semantic segmentation models, providing comprehensive performance benchmarks and comparative analysis. This study offers essential data support and an evaluation benchmark for visual detection and intelligent monitoring applications, with the potential to advance the development of ethylene leakage segmentation algorithms.