Deep learning method and apparatus for automatic upright rectification of virtual reality content
Proposed are a deep learning method and apparatus for the automatic upright rectification of VR content. The deep learning method for the automatic upright rectification of VR content according to an embodiment may include inputting a VR image, to a neural network and outputting orientation information of the VR image through a trained neural network.
1. A deep learning method for automatic upright rectification of VR (Virtual Reality) content, comprising:
inputting a VR image to a neural network;
automatically generating a dataset of VR images composed of numerous VR images having ground truth orientations in order to train the neural network; and
outputting orientation information (up-vector) of the input VR image through a trained neural network,
wherein automatically generating the dataset of VR images comprises:
automatically downloading numerous VR images from the Internet;
assuming these VR images to be an upright state;
generating new VR images by synthetically rotating these VR images to obtain various orientations; and
training the neural network using the generated dataset of VR images,
the method further comprising:
upright-rectifying the VR image by applying a rotation that maps the output up-vector to the straight direction (0,0,1) to the VR image, and
evaluating an amount of mis-orientation of the upright-rectified VR image based on the orientation information (up-vector) output through the deep learning neural network,
wherein inputting the VR image, to the neural network comprises rotating the VR image in a plurality of orientations and inputting this plurality of rotated VR images to the deep learning neural network, and
wherein the evaluating the amount of mis-orientation of the upright-rectified VR image comprises estimating an evaluation metric of the upright-rectified VR images through a deep learning neural network-based evaluator and then returning a VR image having a small tilted degree.
2. The deep learning method of claim 1 , wherein inputting the VR image to the deep learning neural network comprises inputting the VR image to a convolutional neural network (CNN).
3. The deep learning method of claim 1 , wherein the orientation information has been parameterized as at least one of a unit vector, rotation angles, azimuth+elevation angles, and quaternion,
wherein these parametrizations define an up-vector, corresponding to the orientation of the input image.