object-centric grasp

[JudyYe]

Hi,

Nice work! Thank you for releasing data and code! As far as I understand, there are 3 sets of grasps. 1) 367 grasps from manual annotation. (sec3.1 para 1) 2) many more object-centric grasps generated by symmetry. (sec3.1 para 2) 3) transfer 2) to YCB-Video dataset after filtering collision (28M).

I notice that set 1 is in data/grasps, set 3 in YCB_Affordance_grasps. I wonder if set 2 is released. If not, could you please kindly provide the code of generation procedure from set1 to set2?

Thank you very much!

[enriccorona]

Hi,

Good point, this file contains the code to generate object-centric grasps after symmetry, and the generated grasps (set 2) of the final version of the dataset: symmetry_code.zip

Set 2 contains a bit more than 800 grasps but you can create many more depending on a few hyperparameters. Actually, I thought we had more of them after this step. I haven't used the code for a while but I think it should work when setting correct paths.

Best, Enric

[JudyYe]

Thank you for your quick reponse. I will run it today.

[JudyYe]

Hi,

Thank you for providing symmetry code and data. However, could you please explain how to interpret / visualize the extended grasp?

I realize that you assign a new mano_trans and pca_manorot, which are used in visualize_grasps.py. However, when I visualize it in the same way as that for set1. I got the following figures:

(visualizing grasp_0_3.pkl)

I also notice that the core transformation is in Line 74-75.

    newpos = np.matmul(position, apply_rotmat)
    newrot = np.matmul(hand_rotmat, apply_rotmat)

But I don't quite understand it. Suppose positions are column vector (which seems true for MANO layer) ,where S is apply_rotmat. I would assume the transformation just apply an additional rotation around object rotation axis (on the left). Mabye something like , though it's not the exact form since rotational axis and MANO origin offset should also be considered.

I wonder if I misunderstood at some point and how to correctly visualize the generated grasps. Looking forward to your reply. Thank you!