Thank you for releasing the ABC-130K dataset.
Regarding camera extrinsic calibration, we found that the wrist-camera extrinsics can be recovered through forward kinematics, while the head-camera extrinsics can be approximately estimated using SLAM.
However, we observed a large number of different extrinsic configurations across the dataset. We suspect that the data may have been collected using multiple robot platforms or data-collection setups.
Although it is possible to estimate the head-camera extrinsics independently for each episode using SLAM, doing so for every episode would be computationally expensive and inefficient.
Would it be possible to provide a mapping between episode IDs and the corresponding robot platform or data-collection rig? With this information, we could estimate the extrinsics once for each platform and reuse them across episodes collected with the same setup, which would significantly improve the efficiency of extrinsic recovery.
Thank you again for making this dataset available.
Thank you for releasing the ABC-130K dataset.
Regarding camera extrinsic calibration, we found that the wrist-camera extrinsics can be recovered through forward kinematics, while the head-camera extrinsics can be approximately estimated using SLAM.
However, we observed a large number of different extrinsic configurations across the dataset. We suspect that the data may have been collected using multiple robot platforms or data-collection setups.
Although it is possible to estimate the head-camera extrinsics independently for each episode using SLAM, doing so for every episode would be computationally expensive and inefficient.
Would it be possible to provide a mapping between episode IDs and the corresponding robot platform or data-collection rig? With this information, we could estimate the extrinsics once for each platform and reuse them across episodes collected with the same setup, which would significantly improve the efficiency of extrinsic recovery.
Thank you again for making this dataset available.