UTLD: An Underground Thermal and LiDAR Dataset for Depth Estimation

• Introduction

Depth estimation is a crucial task in photogrammetry and computer vision. The underground scenes, characterized by low-light conditions, high dusty, and narrow structures, pose challenges in depth estimation using existing visual-based datasets. We provide an Underground Thermal image and Lidar Dataset (UTLD) for depth estimation over underground scenes. It contains stereo thermal images and the corresponding point clouds achieved by stereo laser scanners over three different underground mines. We tested some monocular depth estimation methods on the UTLD dataset to highlight the challenges and opportunities. Figures 1-2 show the acquisition scenes and platforms, respectively.

Figure. 1. UTLD dataset real collection environment

Figure. 2. Data Collection Platform

• Method Testing

We selected four existing monocular depth estimation methods, each implemented using their official source codes. Figure 3 compares the depth maps of different mathods on the dataset. The methods predict large objects well but struggle with distant targets and fine-grained details. Nevertheless, they capture the geometric structures. Besides, we presents the evaluation metrics for these methods on the UTLD dataset, where the PixelFormer method achieves the best performances (not included in the text).

   

Figure. 3. Depth maps of different methods on the UTLD dataset.

• Conclusion & Prospects

This study introduces the UTLD dataset and validates the feasibility of monocular depth estimation methods in underground mines. In future work, we will improve the image quality under high dust underground scenes. Besides, semantic segmentation will be involved to promote the practical adoption of vision systems in smart applications of underground mines.