Rolling shutter rectification in images/videos using convolutional neural networks with applications to SFM/SLAM with rolling shutter images/videos
A method for correcting rolling shutter (RS) effects is presented. The method includes generating a plurality of images from a camera, synthesizing RS images from global shutter (GS) counterparts to generate training data to train the structure-and-motion-aware convolutional neural network (CNN), and predicting an RS camera motion and an RS depth map from a single RS image by employing a structure-and-motion-aware CNN to remove RS distortions from the single RS image.
1. A computer-implemented method executed on a processor for correcting rolling shutter (RS) effects, the method comprising:
generating a plurality of images from a camera;
synthesizing RS images from global shutter (GS) counterparts to generate training data to train the structure-and-motion-aware convolutional neural network (CNN); and
predicting an RS camera motion and an RS depth map from a single RS image by employing a structure-and-motion-aware CNN to remove RS distortions from the single RS image.
2. The method of claim 1 , further comprising, during a training phase, employing a set of GS images and corresponding GS depth maps to generate synthesized RS camera motions and synthesized RS depth maps.
3. The method of claim 2 , further comprising providing the synthesized RS camera motions, the synthesized RS depth maps, and the synthesized RS images to the structure-and-motion-aware CNN as supervision signals and input data, respectively, to train the structure-and-motion-aware CNN.
4. The method of claim 3 , further comprising, during a testing phase, providing the single RS image to the structure-and-motion-aware CNN to predict the RS camera motion and the RS depth map corresponding to the input RS image.
5. The method of claim 4 , further comprising transmitting the single RS image, the RS camera motion, and the RS depth map to a rectification module to remove the RS effects.
6. The method of claim 5 , further comprising outputting a rectified image from the rectification module.
7. The method of claim 1 , wherein the structure-and-motion-aware CNN includes a depth network for learning the RS depth map and a velocity network for learning the RS camera motion.
8. The method of claim 7 , wherein the depth network is trained separately from the velocity network.
9. A non-transitory computer-readable storage medium comprising a computer-readable program for correcting rolling shutter (RS) effects, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:
generating a plurality of images from a camera;
synthesizing RS images from global shutter (GS) counterparts to generate training data to train the structure-and-motion-aware convolutional neural network (CNN); and
predicting an RS camera motion and an RS depth map from a single RS image by employing a structure-and-motion-aware CNN to remove RS distortions from the single RS image.
10. The non-transitory computer-readable storage medium of claim 9 , wherein, during a training phase, a set of GS images and corresponding GS depth maps are employed to generate synthesized RS camera motions and synthesized RS depth maps.
11. The non-transitory computer-readable storage medium of claim 10 , wherein the synthesized RS camera motions, the synthesized RS depth maps, and the synthesized RS images are provided to the structure-and-motion-aware CNN as supervision signals and input data, respectively, to train the structure-and-motion-aware CNN.
12. The non-transitory computer-readable storage medium of claim 11 , wherein, during a testing phase, the single RS image is provided to the structure-and-motion-aware CNN to predict the RS camera motion and the RS depth map corresponding to the input RS image.
13. The non-transitory computer-readable storage medium of claim 12 , wherein the single RS image, the RS camera motion, and the RS depth map are transmitted to a rectification module to remove the RS effects.
14. The non-transitory computer-readable storage medium of claim 13 , wherein a rectified image is output from the rectification module.
15. The non-transitory computer-readable storage medium of claim 9 , wherein the structure-and-motion-aware CNN includes a depth network for learning the RS depth map and a velocity network for learning the RS camera motion.
16. The transitory computer-readable storage medium of claim 15 , wherein the depth network is trained separately from the velocity network.
17. A system for correcting rolling shutter (RS) effects, the system comprising:
a memory; and
one or more processors in communication with the memory configured to:
generate a plurality of images from a camera;
synthesize RS images from global shutter (GS) counterparts to generate training data to train the structure-and-motion-aware convolutional neural network (CNN); and
predict an RS camera motion and an RS depth map from a single RS image by employing a structure-and-motion-aware CNN to remove RS distortions from the single RS image.
18. The system of claim 17 , wherein, during a training phase, a set of GS images and corresponding GS depth maps are employed to generate synthesized RS camera motions and synthesized RS depth maps.
19. The system of claim 18 , wherein the synthesized RS camera motions, the synthesized RS depth maps, and the synthesized RS images are provided to the structure-and-motion-aware CNN as supervision signals and input data, respectively, to train the structure-and-motion-aware CNN.
20. The system of claim 19 , wherein, during a testing phase, the single RS image is provided to the structure-and-motion-aware CNN to predict the RS camera motion and the RS depth map corresponding to the input RS image.