maximiliangilles / MetaGraspNet

MetaGraspNet: Official Code Repository
Other
87 stars 10 forks source link

Difficulty Level #12

Open StephanyChanelo opened 6 months ago

StephanyChanelo commented 6 months ago

Dear authors,

I was reading the MetaGraspNetV2 paper, in particular the section V.B Task-Specific Difficulty Levels. When I was trying to use the same approach to calculate the difficulty per scene, I came to realize that in the paper is not clear how these attributes are calculated or where they are obtained.

mgnv2

  1. Is this information included on the dataset?
  2. Is the difficulty score storage in some file?
maximiliangilles commented 6 months ago

Hi @StephanyChanelo ,

I've uploaded both the vacuum scene difficulty splits and the individual difficulty scores for each scene (air-permeability, gripper dimension, shape, specularity, texturedness, occlusion, and transparency) for the MGNv2-Real dataset to our shared folder. As detailed, the vacuum scene difficulty score is calculated by summing up these individual scores. We've applied K-Means clustering to categorize the scenes into three difficulty levels: easy, medium, and hard. Depending on your specific needs, you may find some difficulty scores more relevant than others. It might be worth adjusting these scores to better fit your use case, see $\lambda_p$ in EQ(1).

I hope this information is helpful to you.

StephanyChanelo commented 6 months ago

Hi @StephanyChanelo ,

I've uploaded both the vacuum scene difficulty splits and the individual difficulty scores for each scene (air-permeability, gripper dimension, shape, specularity, texturedness, occlusion, and transparency) for the MGNv2-Real dataset to our shared folder. As detailed, the vacuum scene difficulty score is calculated by summing up these individual scores. We've applied K-Means clustering to categorize the scenes into three difficulty levels: easy, medium, and hard. Depending on your specific needs, you may find some difficulty scores more relevant than others. It might be worth adjusting these scores to better fit your use case, see λp in EQ(1).

I hope this information is helpful to you.

Thank you for the information, we are wondering if this information is available also for the synthetic dataset in some file?

maximiliangilles commented 6 months ago

Hi @StephanyChanelo,

Unfortunately, we do not provide these difficulties for the synthetic data. We focused on the real-world dataset because real-world objects pose unique challenges to vision-based grasping, e.g. transparency, specularity, or air permeability.

Best,

Maximilian