LoliMVS: An End-to-End Network for Multiview Stereo With Low-Light Images
Yangang Wang, Qingfang Jiang · IEEE Transactions on Instrumentation and Measurement · 2024
Performing multiview stereo (MVS) reconstruction under a low-light environment is challenging. Different from traditional MVS methods that work with images captured under normal lighting conditions, we focus on reconstructing 3-D models in low-light situations. To address this, we propose a learning-based MVS framework consisting of two steps: low-light image enhancement and MVS reconstruction. At first, we adopt an encoder-decoder network to enhance the low-light images, making them more visually discernible. Then, the encoder is fixed and shared, facilitating to train a decoder for MVS reconstruction. To validate our approach, we have created a new dataset called LoLi100, specifically designed for low-light reconstruction. In our experiments, we train and test our method on this dataset, demonstrating that our pipeline generates 3-D models with fine details and clear texture. Compared to other methods, our approach significantly improves completeness and overall quality of the depth maps. The dataset is publicly available athttps://www.yangangwang.com.