korrawe / grasping_field Goto Github PK

Dear Authors,

Really nice work! Thank you for making the code available.
Would you be releasing the pre-trained model for RGB to 3D reconstruction network?

Thanks!

Dear Author,

Thanks for making the code available. I see that you have used obMan dataset for training, as far as I have understood obMan object meshes are non-watertight and contain very thin structures which cannot be closed. I am wondering how did you manage to compute signed distance fields/occupancy labels?

It would be really helpful if you could provide some insight on how did you process the obMan dataset and also how the labels are computed for implicit function representation?

Thanks in Advance!

Hi, thanks for your great work.

There are some questions I feel confused.

First, you said' New objects can be given to the model by providing the path to the meshes (./input) and the list of object (input.json). The object needs to be reachable when a hand wrist is at the origin and should not be in [-x,-z] quadrant (see example meshes).' I am not sure whether it means I need make all of vertices of objects (x>0,z>0)?
Second, about the scale of object? I download some objs from YCB-benchmark. I find the scale of each obejct is different. So whether I need scale the object to fixed size?
Finally, questions about the number of vertices of objs? could you tell me the range of the number of vertices of object is suitable to your codes?

Look forward to your reply.

best
Qiu

Hi,
Thank you for your great job. We are studying the hand pose estimation and want to compare the results on ObMan with your method. So we wish to get the PCK and AUC of your method on ObMan such that we can compare with your method. Can you offer them?
Thank you!

The pre-processing procedure of DeepSDF seems to be very challenging and contain some obstacles. Could you release the pre-processed SDF data used for the Grasping Field?

This file is not included in obman or ShapeNetCoreV2. How to find it?

Hi korrawe. Thanks for sharing this nice work. I am very interested in this repo, but I find the training procedure is hard to follow. Could you release more detailed instructions, such as how to prepare obman data, how to select training parameters?
Looking for your reply. Thanks in advance.

Hello!

I'm very interested in your test results and I'm trying to implement it.

However, I am having some difficulties and I have a question!

How did you handle data in HO-3D or FPHA in the point cloud setting?

More specifically, I know I need to pass the model after moving the HO-3D or FPHA into obman object space, how did you do this?

I will be waiting for your or any others reply.

Thank you! :)