Dear Mr. Stachniss, Thank you very much for your lecture. I am a PhD student working on 3D reconstruction / photogrammetry and sensor data fusion in Nürnberg. Your lectures, which you are publishing online for open access available for everyone, are outstanding! You are providing a great way for people to learn from your experience in an easy to understand didactic approach with logical examples. I have bought the "Photogrammetric Computer Vision" book and am currently trying to establish an Evaluation environment for Visual SLAM research (github.com/GSORF/Visual-GPS-SLAM) - your lectures will be very useful on my journey. Thank you very much and all the best for you! Grateful regards!
I missed the Part#1......When I saw this Video, came to know there was a Part#1......Will first catch-up Part#1 and view and take notes of these important topics.......Thanks a lot Professor......
Great lecture! 2 questions: 1) why N=At.Sigma⁻1.A where Sigma looks like covariance matrix (and so Sigma^-1 is information matrix)? Why not N=At.Sigma.A? 2) Could you give some pointers/reference on how to compute/estimate the jacobian A?
At 14:02 I am not sure what that 2 in, 2x171=342 observation represents. Could someone clarify? Does it mean 2 cameras, or is it because its the x and y dimensions?
I am confused by the block sizes in the C matrix in the graphic illustration. Could you clarify the meaning/sizes of the black squares in the C matrix and why they seem to appear in formations of 2 and 3 diagonally? Thank you!
Each black rectangle is the C_{ij} matrix of size 2x3. It means that the 3d point "i" was observed by the camera/image "j". The top six rectangular blobs in C correspond to points 1, 2, 8, 9, 15, 16 which were observed in the first image. Since these points are close by, they actually look like 3 blocks instead of 6. I believe the nice diagonal structure is due to the flight pattern (24:20) and (26:20)
Hello Cyrill. You know when you stacked the camera orientation parameters for Δt, you included only the extrinsic parameters. If you want to estimate the intrinsic parameters as well (i.e. bundle adjustment for the uncalibrated case) do you also stack those 5 intrinsic variables with the 6 extrinsic parameters and make it into a size 11 vector and the C_{ij} ending up becoming 2x11 blocks? Very appreciate it if you can clarify this. Otherwise I very much enjoyed your lecture!
Sure - bot only if you want the 6 intrinsics to be image-specific. If you use the same camera for recording the images, you want 6 intrinsics holding for all images.
I think the vector h could be computed by A^T \Sigma^{-1} \Delta l, which can be found ( czcams.com/video/LKDLcKrWOIU/video.html ). And hk, ht are just two blocks of vector h, just like how we divide matrix A to C, B blocks
These lectures are the best resource I found so far in SLAM topics, Thank you Sir
Dear Mr. Stachniss,
Thank you very much for your lecture. I am a PhD student working on 3D reconstruction / photogrammetry and sensor data fusion in Nürnberg. Your lectures, which you are publishing online for open access available for everyone, are outstanding! You are providing a great way for people to learn from your experience in an easy to understand didactic approach with logical examples. I have bought the "Photogrammetric Computer Vision" book and am currently trying to establish an Evaluation environment for Visual SLAM research (github.com/GSORF/Visual-GPS-SLAM) - your lectures will be very useful on my journey. Thank you very much and all the best for you!
Grateful regards!
Great, I am happy to hear that and thanks for the link!
I missed the Part#1......When I saw this Video, came to know there was a Part#1......Will first catch-up Part#1 and view and take notes of these important topics.......Thanks a lot Professor......
Thank you very much, Professor! It helped me to understand much better the problem. Subscribed and looking forward to learn more from you!
Great lecture! 2 questions: 1) why N=At.Sigma⁻1.A where Sigma looks like covariance matrix (and so Sigma^-1 is information matrix)? Why not N=At.Sigma.A? 2) Could you give some pointers/reference on how to compute/estimate the jacobian A?
At 14:02 I am not sure what that 2 in, 2x171=342 observation represents. Could someone clarify? Does it mean 2 cameras, or is it because its the x and y dimensions?
It is basically x and y coordinates of points in the image plain. 2 observations per point.
Correct
I am confused by the block sizes in the C matrix in the graphic illustration. Could you clarify the meaning/sizes of the black squares in the C matrix and why they seem to appear in formations of 2 and 3 diagonally? Thank you!
Each black rectangle is the C_{ij} matrix of size 2x3. It means that the 3d point "i" was observed by the camera/image "j". The top six rectangular blobs in C correspond to points 1, 2, 8, 9, 15, 16 which were observed in the first image. Since these points are close by, they actually look like 3 blocks instead of 6. I believe the nice diagonal structure is due to the flight pattern (24:20) and (26:20)
Hello Cyrill. You know when you stacked the camera orientation parameters for Δt, you included only the extrinsic parameters. If you want to estimate the intrinsic parameters as well (i.e. bundle adjustment for the uncalibrated case) do you also stack those 5 intrinsic variables with the 6 extrinsic parameters and make it into a size 11 vector and the C_{ij} ending up becoming 2x11 blocks?
Very appreciate it if you can clarify this.
Otherwise I very much enjoyed your lecture!
Sure - bot only if you want the 6 intrinsics to be image-specific. If you use the same camera for recording the images, you want 6 intrinsics holding for all images.
@@CyrillStachniss , thank you :)
Very nice lecture. Just one small question, what is hk and ht and how do we obtain them?
I think the vector h could be computed by A^T \Sigma^{-1} \Delta l, which can be found ( czcams.com/video/LKDLcKrWOIU/video.html ). And hk, ht are just two blocks of vector h, just like how we divide matrix A to C, B blocks
Thank you so much.
Thank you professor, Can MATLAB be very ok for this operations?
Using sparse matrices, yes.
Can we get the slides for these lectures
Send me an email