Hello, Kevin, I would like to perform real-time 3D rotation estimation of a cuboid (with rotations along the X, Y, and Z axes not exceeding 30 degrees). Would it be feasible to use SIFT to manually annotate four points on one face of the cuboid and then use CNN for convolutional training to accomplish this task? How would this approach perform in terms of accuracy and real-time capability? Alternatively, do you have any other ideas for solving this problem? I would greatly appreciate your input.
In theory that could work. But I’d say the easiest way is to use some markers if you can attach something on the cube or checkout my latest video on 6D pose estimation
@@kevinwoodrobotics Sorry, Kevin, I was busy with other things and missed your RUclips message from a few days ago. When you mentioned adding markers to objects, are you referring to something like adding a QR code to the object, similar to AirTags, or something like the codes in your new video? If you don't add markers, does that mean using RGB or RGBD images in CNN training might result in lower real-time performance, making it less suitable for real-time detection and correction environments? Looking forward to your reply.
Great question! DOPE can be good when it’s difficult to find specific points to detect your pose, especially organic objects. But if you can reliably figure out key points, then you can use this classical approach
Generally yes if the points provide new information. If the points are identical or if they are clustered or same plane, it may cause problems or have a degenerate solution
Hello May I ask Why i got this error? [Errno 2] No such file or directory: 'C:\\Users\\OneDrive - \\Documents\\Pose Estimation\\54-camera-calibration/calibration.npz' File "C:\Users\OneDrive - \Documents\Pose Estimation\main.py", line 38, in poseEstimation data = np.load(paramPath) ^^^^^^^^^^^^^^^^^^ File "C:\Users\OneDrive - \Documents\Pose Estimation\main.py", line 81, in poseEstimation(DrawOption.CUBE) FileNotFoundError: [Errno 2] No such file or directory: 'C:\\Users\\OneDrive - \\Documents\\Pose Estimation\\54-camera-calibration/calibration.npz' I just copied and pasted the code in your github in my vs code, is there any configurations I need to do? thank you.
Code and Doc: kevinwoodrobotics.com/product/opencv-python-pose-estimation/
OpenCV Python Playlist Code and Doc: kevinwoodrobotics.com/product/opencv-python-tutorials-full-playlist/
Hello, Kevin, I would like to perform real-time 3D rotation estimation of a cuboid (with rotations along the X, Y, and Z axes not exceeding 30 degrees). Would it be feasible to use SIFT to manually annotate four points on one face of the cuboid and then use CNN for convolutional training to accomplish this task? How would this approach perform in terms of accuracy and real-time capability? Alternatively, do you have any other ideas for solving this problem? I would greatly appreciate your input.
In theory that could work. But I’d say the easiest way is to use some markers if you can attach something on the cube or checkout my latest video on 6D pose estimation
For real time there’s a lot of variables to consider so hard to say without more details
@@kevinwoodrobotics Sorry, Kevin, I was busy with other things and missed your RUclips message from a few days ago. When you mentioned adding markers to objects, are you referring to something like adding a QR code to the object, similar to AirTags, or something like the codes in your new video? If you don't add markers, does that mean using RGB or RGBD images in CNN training might result in lower real-time performance, making it less suitable for real-time detection and correction environments? Looking forward to your reply.
Hi great video! What is the downside of this implementation compared to 6D pose estimation by training 3D object with DOPE?
Great question! DOPE can be good when it’s difficult to find specific points to detect your pose, especially organic objects. But if you can reliably figure out key points, then you can use this classical approach
@@kevinwoodrobotics did you make a video about DOPE implementation?
@@patboy24 I didn’t but I’ve looked into it a bit
Hi, cool video!
What is the version of opencv are you using here?
opencv-python==4.8.0.74
@@kevinwoodrobotics thanks a lot)
@@kevinwoodrobotics btw will you make a video about Aruco markers detection?
@@densq31 sure I can add to my list!
Hi will the accuracy of the pose increase if I use more than 4 points in sovepnp?
Generally yes if the points provide new information. If the points are identical or if they are clustered or same plane, it may cause problems or have a degenerate solution
Hello May I ask Why i got this error?
[Errno 2] No such file or directory: 'C:\\Users\\OneDrive - \\Documents\\Pose Estimation\\54-camera-calibration/calibration.npz'
File "C:\Users\OneDrive - \Documents\Pose Estimation\main.py", line 38, in poseEstimation
data = np.load(paramPath)
^^^^^^^^^^^^^^^^^^
File "C:\Users\OneDrive - \Documents\Pose Estimation\main.py", line 81, in
poseEstimation(DrawOption.CUBE)
FileNotFoundError: [Errno 2] No such file or directory: 'C:\\Users\\OneDrive - \\Documents\\Pose Estimation\\54-camera-calibration/calibration.npz'
I just copied and pasted the code in your github in my vs code, is there any configurations I need to do? thank you.
It could be where you’re running it from or the structure of your folders. Can try checking that or using some manual paths of all else fails
@@kevinwoodrobotics yep I just made a file in VS code then made 2 files inside that one for pose stimation and 1 for camera callibration