Abstract: Disclosed herein is an image deep learning model training method. The method includes sampling a twin negative comprising a first negative sample and a second negative sample by selecting the first negative sample with a highest similarity out of an anchor sample and a positive sample constituting a matching pair in each class and by selecting the second negative sample with a highest similarity to the first negative sample, and training the samples to minimize a loss of a loss function in each class by utilizing the anchor sample, the positive sample, the first and second negative samples for each class. The first negative sample is selected in a different class from a class comprising the matching pair, and the second negative sample is selected in a different class from classes comprising the matching pair and the first negative sample.

Abstract: Disclosed herein are an apparatus and method for detecting a keypoint based on deep learning robust to scale changes based on information change across receptive fields. The apparatus for detecting a keypoint based on deep learning robust to scale changes based on information change across receptive fields includes a feature extractor for extracting a feature from an input image based on a pre-trained deep learning neural network, an information accumulation pyramid module for outputting, from the feature, at least two filter responses corresponding to receptive fields having different scales, an information change detection module for calculating an information change between the at least two filter responses, a keypoint detection module for creating a score map having a keypoint probability of each pixel based on the information change, and a continuous scale estimation module for estimating a scale of a receptive field having a biggest information change for each pixel.

Abstract: Disclosed herein is an image deep learning model training method. The method includes sampling a twin negative comprising a first negative sample and a second negative sample by selecting the first negative sample with a highest similarity out of an anchor sample and a positive sample constituting a matching pair in each class and by selecting the second negative sample with a highest similarity to the first negative sample, and training the samples to minimize a loss of a loss function in each class by utilizing the anchor sample, the positive sample, the first and second negative samples for each class. The first negative sample is selected in a different class from a class comprising the matching pair, and the second negative sample is selected in a different class from classes comprising the matching pair and the first negative sample.

Abstract: An apparatus and a method for displaying a panoramic image using a look-up table (LUT) are disclosed, including generating an LUT may include determining first geometric correction information to transform an input domain pixel coordinate system of an input image to a panorama domain pixel coordinate system of a panoramic image, determining second geometric correction information to transform an output domain pixel coordinate system of an output image of the panoramic image to the panorama domain pixel coordinate system of the panoramic image, determining third geometric correction information to transform the output domain pixel coordinate system of the output image to the input domain pixel coordinate system of the input image based on the first geometric correction information and second geometric correction information, and generating an LUT that maps the output domain pixel coordinate system of the output image to the input domain pixel coordinate system of the input image.

Abstract: An apparatus and a method for displaying a panoramic image using a look-up table (LUT) are disclosed, including generating an LUT may include determining first geometric correction information to transform an input domain pixel coordinate system of an input image to a panorama domain pixel coordinate system of a panoramic image, determining second geometric correction information to transform an output domain pixel coordinate system of an output image of the panoramic image to the panorama domain pixel coordinate system of the panoramic image, determining third geometric correction information to transform the output domain pixel coordinate system of the output image to the input domain pixel coordinate system of the input image based on the first geometric correction information and second geometric correction information, and generating an LUT that maps the output domain pixel coordinate system of the output image to the input domain pixel coordinate system of the input image.

Abstract: Provided is a geometric correction apparatus and method based on a recursive Bezier patch sub-division. A geometric correction method may include: receiving, from a camera, a first image that is obtained by photographing a black screen that is projected by a projector onto a projection surface; receiving, from the camera, a second image that is obtained by photographing a predetermined pattern that is projected by the projector onto the projection screen; generating a third image by subtracting the first image from the second image; and performing geometric correction with respect to the predetermined pattern to correct a distortion between the predetermined pattern and the third image.

Abstract: An apparatus and a method for correcting colors of an image projection device are provided. The method includes: acquiring a photographed image by photographing a sample image projected on projection surface; generating input-output color information for n regions, based on color values of a block in the sample image and corresponding color values of the block in the photographed image; selecting one of the n regions of photographed images as a reference region; generating look-up tables (LUTs) for non-reference regions, based on the reference region and the input and output color information; and correcting colors of input images to be projected by the image projection device using the look-up tables, thereby minimizing color difference of the input images on the projection surface for both intra and inter projection device color correction while simplifying the correction procedure.

Abstract: The present invention proposes a method and apparatus for correcting a motion of panorama video captured by a plurality of cameras. The method of the present invention includes performing global motion estimation for estimating smooth motion trajectories from the panorama video, performing global motion correction for correcting a motion in each frame of the estimated smooth motion trajectories, performing local motion correction for correcting a motion of each of the plurality of cameras for the results in which the motions have been corrected, and performing warping on the results on which the local motion correction has been performed.

Abstract: Disclosed are a color correction apparatus for panorama video stitching and a method of selecting a reference image using the same. A method of selecting a reference image for color correction when stitching panorama video based on input images includes selecting an optimum reference image candidate from the input images based on standard deviations for overlapping regions between the input images, performing color correction on the input images based on the optimum reference image candidate, and validating the optimum reference image candidate based on the color-corrected input images.