Monocular Depth Estimation with Convolutional Neural Networks on Embedded Systems
Monocular depth estimation is becoming a very interesting problem in computer vision to solve due to the several tasks that require as an input the spatial structure of a scene, such as 3D reconstruction, 3D object detection, localization and mapping. The most effective techniques for monocular dept...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| País: | México |
| Institución: | Universidad Autónoma de Querétaro |
| Repositorio: | Repositorio Institucional de la Universidad Autónoma de Querétaro |
| Idioma: | inglés |
| OAI Identifier: | oai:ri-ng.uaq.mx:123456789/3397 |
| Acceso en línea: | http://ri-ng.uaq.mx/handle/123456789/3397 |
| Access Level: | acceso abierto |
| Palabra clave: | monocular depth low-latency convolutional OTRAS |
| Sumario: | Monocular depth estimation is becoming a very interesting problem in computer vision to solve due to the several tasks that require as an input the spatial structure of a scene, such as 3D reconstruction, 3D object detection, localization and mapping. The most effective techniques for monocular depth estimation are based on large deep learning-based architectures that cannot be deployed on systems with limited computational resources and therefore preventing its use in application fields where the advantages of monocular cameras (i.e., low cost, small size, low weight and low-energy consumption) could also be exploited. Under this context, the research of low-latency deep learning architectures for monocular depth estimation is a very promising topic for which just a few methods have been proposed until now. In this master thesis, a very low-latency fully convolutional network is proposed. The quantitative results on the NYU-Depth V2 dataset show that the proposed method is 1.6x faster than the state-of-the art related method while also reducing the RMSE metric by 1.16%. |
|---|