Hyperparameter Optimization using Grid Search for use in Monocular Depth Estimation

Autonomous vehicles uses a myriad of sensors to observe and monitor its surroundings. Current re- search incorporates static digital images that are used to train and optimize deep neural networks. These deep neural networks are used to identify and determine the distances of objects from the autonomous vehicle. Unfortunately, the hyperparameters used to train these networks were chosen arbitrarily. We explored the hyperparameter space through a two-step grid search in order to recognize the optimal combination of losses before continuing to train the rest of the system to understand the trade offs be- tween constant parameters. Specifically, the research focuses on finding the optimal relation between the left-right consistency, appearance, disparity losses in generating accurate feature edges in our output images. The trade off between costs allows for the neural network an increased performance at edge detection.