| Meno: | Martin
|
|---|
| Priezvisko: | Melicherčík
|
|---|
| Názov: | Super Sampling of 3D Scans using Deep Learning
|
|---|
| Vedúci: | doc. RNDr. Martin Madaras, PhD.
|
|---|
| Rok: | 2023
|
|---|
| Kµúčové slová: | Super Sampling, Deep Learning, 3D Scans, DLSS
|
|---|
| Abstrakt: | In this thesis, we work on increasing the resolution of depth maps obtained from a 3D camera based on structured light. The aim of the work is to create an output depth map with a higher resolution from a depth map with a lower resolution at the input. In addition to the input depth map, we also have a high-resolution intensity texture of the scene. The depth map super-resolution task is familiar, and several solutions are available based on deep learning models. Deep learning models use intensity textures as auxiliary information. Two models consisting of convolutional neural networks with which we work are FDSR and DKN. We deal with the pre-processing of data from the 3D camera we have chosen so that we can properly train and use the mentioned models. We adjust both models to meet our hardware limitations and make the training process stable. After the data pre-processing and applying the modified trained models, we evaluate the resulting depth maps using qualitative and quantitative metrics focused on the quality of the resulting 3D image. We also evaluate the time efficiency of our solution and its usability.
|
|---|