Depth estimation of a single image
Shikov, Andrei (2017)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2017053011104
https://urn.fi/URN:NBN:fi:amk-2017053011104
Tiivistelmä
The problem of depth estimation is an important component to understand the geometry of a scene and to navigate in space. More knowledge of the surroundings are bringing improvements in other areas, such as in recognition tasks as well. Many efficient algorithms already cover this problem using calibrated stereo cameras. When working with monocular images, a traditional algorithmic approach is experiencing many troubles, as one image can correspond to multiple geometries.
This study is considering neural networks as an algorithm of a depth estimation running on a web server in a cloud. The proposed solution defines concise, accurate and lagless results from a set of images while executing at the server environment with the support of a GPU accelerator. The model achieves relatively high accuracy in identifying the depth, especially boundaries of closer objects while keeping the network operating speeds on the satisfying level.
This study is considering neural networks as an algorithm of a depth estimation running on a web server in a cloud. The proposed solution defines concise, accurate and lagless results from a set of images while executing at the server environment with the support of a GPU accelerator. The model achieves relatively high accuracy in identifying the depth, especially boundaries of closer objects while keeping the network operating speeds on the satisfying level.