Chenfei Liao, Wensong Wang, Zichen Wen, Xu Zheng, Yiyu Wang, Haocong He, Yuanhuiyi Lyu, Lutao Jiang, Xin Zou, Yuqian Fu, Bin Ren, Linfeng Zhang, Xuming Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Specialized datasets designed for mining scenarios are the essential foundation for the development, operation, and research of intelligent mines. Currently, the available datasets focus primarily on open-pit mines, with a lack of specialized datasets for underground mines. This gap severely hinders the application of intelligent solutions in underground mines. This paper proposes a challenging semantic segmentation dataset focusing on underground mines, named the underground mine scenes semantic segmentation (UMSSS) dataset, which contains 4200 high-quality annotated images and 18 annotated categories. To accurately capture the diversity and complexity of mine environments, we collect data from over ten mines located in various geographical regions. The UMSSS dataset is the first open-source semantic segmentation dataset for underground mines, widely covering varying lighting scenarios and diverse underground objects. The comparative experiments extensively explore the characteristics of the UMSSS dataset, providing a detailed evaluation of various state-of-the-art algorithms on the UMSSS dataset.