IJmond Industrial Smoke Segmentation Dataset

The dataset contains raw and cropped images for industrial smoke segmentation. The task is to identify the pixels in the image that belongs to a part of the industrial smoke emission. The standard way for using this dataset is to cast the task as pixel-level classification. However, one can also go for the polygon-based image segmentation approach. We have in total 900 raw images with 1209 polygons. Among them, 7 raw images have no smoke. Also, we have in total 2074 cropped images, and among them, 1109 images have smoke (and have pixel-level masks).If you use the dataset, please cite this technical report (that describes details about the dataset):Yen-Chia Hsu, Despoina Touska. 2026. IJmond Industrial Smoke Segmentation Dataset. arXiv preprint arXiv:2603.23754. https://arxiv.org/abs/2603.23754