6533b7ddfe1ef96bd1274ed6

RESEARCH PRODUCT

The Usage of Quadtree in Deep Neural Networks to Represent Data For Navigation From a Monocular Camera

subject

description

Depth acquisition represents a key element for navigation tasks. It is, therefore, one of the major research topics in computer vision. Many approaches have been developed to address this problem by constructing the depth from series of images. However, there is a minimal case proposing a prediction from a single image, made possible with the emergence of deep learning approaches. The latter makes it possible to consider a reduction in both hardware and computing time costs, which is beneficial for embedded systems. However, network architecture remains a heavy process requiring a lot of GPU memory. Few approaches have proposed addressing this problem by developing lightweight architectures, allowing real-time execution. We propose to investigate this problem from another angle, consisting in carefully selecting the operations to execute rather than lightening the architecture. It is based on the Quadtree Generating Networks framework which takes advantage of sparse convolutions to only operate necessary operations to generate the quadtree, thus reducing the computational cost. This method, initially developed for semantic segmentation, will be applied in this study to data acquisition problems for navigation. Namely, the image segmentation for obstacle avoidance and the generation of compressed depth maps into quadtree. It will be demonstrated, through a series of experiments, that the quadtree compression allows a significant reduction of the memory requirement with a limited loss of accuracy. The level of compression of the prediction is fully adjustable for the depth estimation, making it adaptable to all situations.