Misalignment when transforming object point cloud into scene using provided poses

[vvduc1803]

Hi, thank you for your great work on GraspNet-1Billion.

When visualizing the dataset, I noticed that the transformed object point cloud does not perfectly align with the corresponding object in the scene point cloud. The misalignment is sometimes significant.

In particular:

• The misalignment appears larger when using the Kinect scenes compared to the RealSense scenes.
• The discrepancy is consistent: the object is shifted compared to its real location in the scene.

My questions:

1. What is the main reason for this misalignment? (e.g., calibration error, difference in coordinate systems, scaling, or others?)
2. Does this misalignment affect training accuracy for models trained on GraspNet? How significant is the effect of such errors on grasp pose learning?

I attach one example figure showing the discrepancy (blue: scene, red: transformed object).

Thank you in advance for your clarification!

[DX3906G]

Hi, I encontered the same problem. When I used the provided object pose to render object mask, there was distinct misalignment between rgb image and the rendered mask.

[DX3906G]

Hi, I encontered the same problem. When I used the provided object pose to render object mask, there was distinct misalignment between rgb image and the rendered mask.

Oh sorry, I made a mistake. I flipped the image but I forgot to change the camera intrinsics accordingly. Now I think the annotated object poses are good.

[vvduc1803]

Oh sorry, I made a mistake. I flipped the image but I forgot to change the camera intrinsics accordingly. Now I think the annotated object poses are good.

Could you show me some image examples of good poses in Kinect camera? Thanks!

[DX3906G]

Oh sorry, I made a mistake. I flipped the image but I forgot to change the camera intrinsics accordingly. Now I think the annotated object poses are good.

Could you show me some image examples of good poses in Kinect camera? Thanks!

object mask projected on rgb image

[vvduc1803]

Oh sorry, I made a mistake. I flipped the image but I forgot to change the camera intrinsics accordingly. Now I think the annotated object poses are good.

Could you show me some image examples of good poses in Kinect camera? Thanks!

object mask projected on rgb image

Can you provide me the code to run this result because i used the poses and code provided but still got wrong result?