Questions regarding inaccuracy of extrinsic parameters in *_extrinsic_params_aligned.json

[hetolin]

Hi, I am using the hand_left_extrinsic_params_aligned.json and head_extrinsic_params_aligned.json files to project the pose data read from /state/end/position to 2D image coordinates. When visualizing the projected axes, they appear on the flange of the robot arm. In order to move the projection to the gripper, I added a 0.227 m offset to the pose (from /state/end/position). However, the visualization suggests that the extrinsic alignment is still not correct.

Pose directly read from /state/end/position:

Add 0.227m offset:

(1) Is there any recommended way to solve or reduce the inaccuracy in the extrinsic parameters provided in the *_extrinsic_params_aligned.json files?

(2) Could you clarify the meaning of the extrinsic matrices in these files?
Does hand_left_extrinsic_params_aligned.json describe the transformation from the hand_left camera frame to which robot frame?
Does head_extrinsic_params_aligned.json describe the transformation from the head camera frame to which frame?
If the extrinsics are not precise, is it advisable to use FK (forward kinematics) with joint angles to compute the transformation from the camera to the base?

Any guidance, clarification, or tools/scripts for recalibrating or recomputing the extrinsic parameters would be greatly appreciated!

Thanks!

[Memories-Apart]

I have the same problem. Some projections are great while others lack accuracy. This makes it difficult to utilize these projection results further.

[hetolin]

The gripper tip projection in your head camera view looks accurate (Fig. 1). Did you include the 0.227 offset or other transformation parameters in this calculation?

Also, have you checked the projection in the wrist camera? There seems to be a noticeable offset from the gripper tip.
@Memories-Apart

[Memories-Apart]

I don't know the exact offset, so I used 0.25 m. However, I think the main problem is the inaccuracy in the directions of the x- and y-axes. I have worked with data from other robotic arms before, so I know it's necessary to add an offset — I tried 0.25 m and it seemed to work well.

And still some(e.g. 603-827752) are out of nowhere...

@hetolin

[hetolin]

Yes, I have the same issue. In some scenes, there’s a large offset in the projection, and some are missing *_extrinsic_params_aligned.json.

[Memories-Apart]

I think the variation inside extrinsic_params_aligned.json is overshadowed by the error of it...
So I just used the one without alignment.

[KaiLiu18]

I don't know the exact offset, so I used 0.25 m. However, I think the main problem is the inaccuracy in the directions of the x- and y-axes. I have worked with data from other robotic arms before, so I know it's necessary to add an offset — I tried 0.25 m and it seemed to work well.

And still some(e.g. 603-827752) are out of nowhere...

[@hetolin](https://github.com/hetolin)

Besides offset, have you performed any other operations on the raw data (such as adding rotation)? If it is possible, could you share the code?

[hetolin]

I only add offset on the raw data @KaiLiu18

[ruoqiyang25]

Hi I wonder where do you find the head_extrinsic_params_aligned.json file as I didn't see it in the parameters directory of the orignal dataset. @hetolin