16. Conclusion


If the correction by object detection and the correction by the previous frame are performed for all 340 frames and the 3D restoration result is displayed while moving the viewpoint, it will look like

Display 340 frames at different angles
. This is the conclusion of this HP for the time being. Stereo vision, stereo parallelism, detection of rotational movement between frames, and object detection have all been done with the gaze_line-depth model and graph cut, but I think that this is all that can be done with the gaze_line-depth model. After that, I think that now I have to combine deep learning. If you are interested, please email oguri.kiyoshi@icloud.com.


Below is a bonus.

3D restored video from 200 continuous stereo images
The original 200 stereo images are
200 consecutive stereo images
. The photo shows me and my laboratory at home. Next is 2011_09_28_drive_0053 downloaded from KITTI Raw Data Person and restored in 3D.
3D restored video from KITTI 2011_09_28_drive_0053 stereo image
The original 200 consecutive stereo images are
Stereo image of KITTI 2011_09_28_drive_0053