Non-line-of-sight imaging from real indirect Time of Flight data

This study delves into the performance of a deep learning model, specifically employing a U-Net architecture, that is trained using synthetic data and evaluated on real-world data. A central achievement of this research is the creation of precise ground truth for the real-world dataset, which plays a pivotal role in evaluating the model’s effectiveness and accuracy. The research primarily addresses the challenges encountered when applying a model, originally trained on highly structured synthetic data, to noisy and limited real-world datasets. To overcome this challenge, a variety of preprocessing techniques, including normalization, are utilized to standardize the intensity ranges and mitigate the noise present in the real-world data. Given the small size of the real dataset, consisting of only 8 images, additional strategies such as data augmentation and transfer learning, with an emphasis on fine-tuning, are employed to enhance the model's ability to generalize effectively. The study also investigates the impact of freezing the encoder during the fine-tuning process, which is designed to improve the model's adaptability to the real-world data while preserving the beneficial features learned from the synthetic data. The findings from this research underscore the limitations imposed by small datasets and reveal the inherent difficulties in transferring models trained on synthetic data to real-world applications. This highlights the necessity for the development of more advanced strategies to bolster model robustness and overall performance. To achieve more reliable and generalized results, the study emphasizes the critical importance of acquiring a larger and more diverse real-world dataset, which would enable the model to perform better across various real-world scenarios.