Depth estimation with stereo thermal and cross modality using deep neural networks

Depth estimation using stereo images is a challenging problem in computer vision. Many different studies have been carried out to solve this problem. With developing machine learning, tackling this problem is often done with neural network-based solutions. FADNet and PSMNet neural networks are two of the leading solutions that have proven their success and continue to be used today. In this thesis, these two studies were used with appropriate modifications in their structures. On the other hand, the images used in these solutions are mostly in the visible spectrum. However, the need to use the Infrared (IR) spectrum for depth estimation has emerged, as it gives beneficial information than the visible spectra in some conditions. In this context, the CATS dataset was found as a usable dataset. However, this dataset could not be used as it is due to the incompatibility of the rectified images available in it. For thermal-thermal image pairs, the raw images in the CATS dataset and spatial point cloud, which are obtained by LiDAR and express the depth of the scene, are made suitable for the training of networks by undergoing various processes. On the other hand, for the visible-thermal image pairs, the existing images in the CATS dataset could not be used due to the difference in the field of view of the visible and thermal cameras, instead, a new dataset was created. In this created dataset, due to the insufficient number of discrete samples obtained from the LiDAR device, sample augmentation was performed. In this study, FADNet and PSMNet neural networks were trained separately with thermal-thermal and visible-thermal image pairs and their results were compared.