search

edwardnguyen1705 / robotic-grasping-cornell

Use CNNs to estimate a grasping point and angle of a given object, so that the robot arm can pick the object.
28 stars 5 forks source link

About cornell dataset #4

Open li-source opened 4 years ago

li-source commented 4 years ago

Hi , how do you splits Cornell dataset on image-wise and object-wise? Thanks!

edwardnguyen1705 commented 4 years ago

Hi @li-source , Actually, I did not split the dataset on object-wise setting. When you use the code to generate data, it is image-wise split. This intuition is that: Image-wise split: training with all objects, while some views remain unseen. This is to evaluate the intra-object generalization capability of the model. Object-wise split: training on all available vies of the same object and testing on new objects. This is to evaluate inter-object generalization capability of the model. Following the above notes, you can split the data whatever method you want. Please read carefully the data description on the web: http://pr.cs.cornell.edu/grasping/rect_data/data.php

li-source commented 4 years ago

Hi @li-source , Actually, I did not split the dataset on object-wise setting. When you use the code to generate data, it is image-wise split. This intuition is that: Image-wise split: training with all objects, while some views remain unseen. This is to evaluate the intra-object generalization capability of the model. Object-wise split: training on all available vies of the same object and testing on new objects. This is to evaluate inter-object generalization capability of the model. Following the above notes, you can split the data whatever method you want. Please read carefully the data description on the web: http://pr.cs.cornell.edu/grasping/rect_data/data.php

thank you for your reply！ and I have an another question. How to generate the bounding box of the object by the target image and the corresponding background image?

edwardnguyen1705 commented 4 years ago

Hi @li-source , Please check ./dataset/*.m/ and the dataloader grasp_dataset.py. I think you can:

  1. Ask @mouad47 in #1. He is working on the same problem.
  2. Check relevant issues here: https://github.com/ivalab/grasp_multiObject_multiGrasp/issues
li-source commented 4 years ago

Hi @li-source , Please check ./dataset/*.m/ and the dataloader grasp_dataset.py. I think you can:

  1. Ask @mouad47 in #1. He is working on the same problem.
  2. Check relevant issues here: https://github.com/ivalab/grasp_multiObject_multiGrasp/issues

ok, thanks!