Synthetic datasets have been used to train 2D and 3D image-based deep learning models, and they serve as also as performance benchmarking. Although some authors already use 3D models for the development of navigation systems, their applications do not consider noise sources, which affects 3D sensors. Time-of-Flight sensors are susceptible to noise and conventional filters have limitations depending on the scenario it will be applied. On the other hand, deep learning filters can be more invariant to changes and take into consideration contextual information to attenuate noise. However, to train a deep learning filter a noiseless ground truth is required, but highly accurate hardware would be need. Synthetic datasets are provided with ground truth data, and similar noise can be applied to it, creating a noisy dataset for a deep learning approach. This research explores the training of a noise removal application using deep learning trained only with the Flying Things synthetic dataset with ground truth data and applying random noise to it. The trained model is validated with the Middlebury dataset which contains real-world data. The research results show that training the deep learning architecture for noise removal with only a synthetic dataset is capable to achieve near state of art performance, and the proposed model is able to process 12bit resolution depth images instead of 8bit images. Future studies will evaluate the algorithm performance regarding real-time noise removal to allow embedded applications.
GPU-Accelerated Time-of-Flight Super-Resolution for Image-Guided Surgery
