SEARCH RESULT

Direction of gravity is a natural way of orienting oneself in an unknown environment. Human beings do thi s wi th equilibrioception. It would be beneficial to estimate the directi on of gravity from a single image for many tasks such as autonomous driving and augmented reality where the knowledge of location of the agent is very important. Extracting this knowledge from an image usually requires a reference to be identified. For example, a traffic light can give away the direction of gravity as it has to be placed in an environment in a specific way with respect to the gravity. Reference-based approaches require a lot of hand modeling for solving the problem. When there are a lot of images with ground truth data is available, one can model these references implici tly using machine learning techniques. We propose to use a set of images along with readings from inertial magnetic unit (IMU) taken from a smart camera observing an indoor environment. This data includes a lot of images as well as ground truth labels for gravity direction extracted from the IMU readings. The data is fed to a convolutional deep neural network to estimate the gravity directions formulated as regression as well as classification problem. We show that this modeling works quite well with a fe w hundred images when we formulate the estimation as a classification problem. The details of the networks trained and the results obtained are presented in depth. Further research with more images but with less accurate ground truth data is underway.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Betül Z. Türkkol Abuzarifa Yakup Genç