search

NVlabs / Deep_Object_Pose

Deep Object Pose Estimation (DOPE) – ROS inference (CoRL 2018)
Other
1.03k stars 287 forks source link

Mix training data from different sources #371

Open joansaurina opened 4 months ago

joansaurina commented 4 months ago

After training for the YCB object mustard, I have obtained results that are not very accurate. After checking the paper, I realized that you were combining two datasets: realistic and randomized.

I have now downloaded the FAT dataset.

Do you recommend training the mustard object of YCB by combining synthetic data from BlenderProc with:

  1. Single images containing only the target object?
  2. A mix of single and mixed images of all objects (so the model also learns not to detect anything)?

Thanks,

Joan

intelligencestreamlabs commented 4 months ago

They have used two approaches for data generation (Blender proc and Nvisii), Nvisii needs linux, nividia drivers and gpus to use it. I have read a tip, you can try to generate image with 5 times of your object and 10 distractors from google_scanned_models

joansaurina commented 4 months ago

Sorry that is not what I'm asking there @intelligencestreamlabs