search

NVlabs / Deep_Object_Pose

Deep Object Pose Estimation (DOPE) – ROS inference (CoRL 2018)
Other
1.02k stars 287 forks source link

Symmetries after data generation #391

Open phsilvarepo opened 1 week ago

phsilvarepo commented 1 week ago

Hello,

I have generated a dataset to train DOPE, in Issac Sim, but there does not seems to exist a method to deal with symmetries, how can I change the model or edit the labels of the generated data to match the requirements to achieve good results?

TontonTremblay commented 1 week ago

Check in this repo, there is information about how to deal with the generated data I think.

phsilvarepo commented 1 week ago

Hello @TontonTremblay,

What I have read so far, it only mentions how to deal with symmetries when generating the data, not after. I already have a dataset, but I cannot achieve good results due to the present symmetries, since pixel-wise, with rotation, the difference will be minimal. To deal with symmetries is the current method, is just to show certain faces during the data generation? Or the method will change the labels?

Thanks

TontonTremblay commented 1 week ago

@mintar do you think he could use the script you wrote to directly modify his dataset? I think so right?

mintar commented 3 days ago

The modifications I made for handling symmetries (https://github.com/NVlabs/Deep_Object_Pose/tree/eb4642d8514071fcf3b2849ca6b8f2ce140b3503/data_generation#handling-objects-with-symmetries) are directly integrated into nvisii_data_gen/single_video_pybullet.py, so this only works when generating a dataset, not for modifying one later.

@phsilvarepo: I'm afraid you'll have to look at the PR where I integrated this (#239) and extract the relevant code and adapt it to modifying your dataset. Sorry!