One of the great teacher in computer vision (Especially kitti dataset handling), Thanks for sharing valuable knowledge. if you have time please share this kind of great stuff in computer vision
Excuse me, I have tried my best but failed to figure out why we should invert the Tmat ,and why T_dot multiplying inversed(Tmat) instead of inversed(Tmat) multiplying T_dot
Couldn't have asked for a better explanation than this. Thanks for putting your thought process out there and sharing your knowledge. Appreciate the effort!!!
Thx for the great series. I'm relatively new to computer vision. Started with cam celebration and depth maps. Now with these videos also a start with odometry. Hope you will start again with new videos if you can find the time.
Thank you for creating the videos and the accompanying notebooks. The depth maps are strangely attractive but as you show not vey accurate and they are expensive to compute. Why not find a set of matching points in the left and right images and then calculate their z coordinate using the formula you provide. Then find the position of the next frame by using solvePNPRansac as before. Note I would use cornersubPixel to try to incresase accuracy. Another approach would be to find the vectors from camera1left to both camera2left AND camera2Right (up to a scale factor) in the camera2 frame. We know that the vector from the camera2left to camera2right is (.54,0,0) and this must be equal to the difference of the direction vectors . So by multiplying each by a scale factor (lambda and gamma say) and subtracting and equating that to (.54,0,0) you can solve for lambda and gamma. Note this is over specified ie there are three equations and two unknowns. I would just use the x and z components but you now have 3 camera positions and perhaps loalized bundle adjustment might be possible. I would love to see how you would approach bundle adjustment. If you are making any more videos could I suggest you have the code full screen and a small version of your ugly face in the top left as I found it difficult to see what you were typing.
Hey, this is good, i need a Dataset for UAVs, i mean, kitti dataset is for a car, do you know a dataset for UAVs ? i need to work with Optical Flow, also i would like to know how to create a dataset
I don't know any open-source drone datasets offhand that have optical flow ground truth, but KITTI does have it. I know you could put an RGBD camera on a drone and make yourself a dataset with depth ground truth pretty easily, but I'm not sure how optical flow ground truth is generated. If you find out, let me know!
Sorry for the late reply. I do want to get to that project, but I'm tied up in some research first before I can get there. Hopefully some more interesting content coming soon!
One of the great teacher in computer vision (Especially kitti dataset handling), Thanks for sharing valuable knowledge. if you have time please share this kind of great stuff in computer vision
Excuse me, I have tried my best but failed to figure out why we should invert the Tmat ,and why T_dot multiplying inversed(Tmat) instead of inversed(Tmat) multiplying T_dot
Couldn't have asked for a better explanation than this. Thanks for putting your thought process out there and sharing your knowledge. Appreciate the effort!!!
Thx for the great series. I'm relatively new to computer vision. Started with cam celebration and depth maps. Now with these videos also a start with odometry. Hope you will start again with new videos if you can find the time.
You saved my life with this tutorial!! Thank you so much!!!
Very nice videos, very nice teacher
Excelent video. I saw the full tutorial, it was very helpful. Thanks a lot.
use a block_size = 7 or 9. Specifically for me, 7 gave better results with fewer missing values (i mean missing value by depth values with over 3800)
Thank you for creating the videos and the accompanying notebooks.
The depth maps are strangely attractive but as you show not vey accurate and they are expensive to compute.
Why not find a set of matching points in the left and right images and then calculate their z coordinate using the formula you provide.
Then find the position of the next frame by using solvePNPRansac as before. Note I would use cornersubPixel to try to incresase accuracy.
Another approach would be to find the vectors from camera1left to both camera2left AND camera2Right (up to a scale factor) in the camera2 frame. We know that the vector from the camera2left to camera2right is (.54,0,0) and this must be equal to the difference of the direction vectors . So by multiplying each by a scale factor (lambda and gamma say) and subtracting and equating that to (.54,0,0) you can solve for lambda and gamma. Note this is over specified ie there are three equations and two unknowns. I would just use the x and z components but you now have 3 camera positions and perhaps loalized bundle adjustment might be possible.
I would love to see how you would approach bundle adjustment. If you are making any more videos could I suggest you have the code full screen and a small version of your ugly face in the top left as I found it difficult to see what you were typing.
Hey, this is good, i need a Dataset for UAVs, i mean, kitti dataset is for a car, do you know a dataset for UAVs ? i need to work with Optical Flow, also i would like to know how to create a dataset
I don't know any open-source drone datasets offhand that have optical flow ground truth, but KITTI does have it. I know you could put an RGBD camera on a drone and make yourself a dataset with depth ground truth pretty easily, but I'm not sure how optical flow ground truth is generated. If you find out, let me know!
@@natecibik5505 are you planning to come up with tutorials any soon?
Hy Nate such nice video. Can you make video about stereo vision on KITTY dataset later ?
hey wont you be doing any vid on kalman filter on this as you said previously? It ill be very helpful ...
Sorry for the late reply. I do want to get to that project, but I'm tied up in some research first before I can get there. Hopefully some more interesting content coming soon!