search

NVIDIA-ISAAC-ROS / isaac_ros_pose_estimation

Deep learned, NVIDIA-accelerated 3D object pose estimation
https://developer.nvidia.com/isaac-ros-gems
Apache License 2.0
142 stars 21 forks source link

AR_data for custom synthetic data made with Isaac Sim Replicator #34

Open monajalal opened 6 months ago

monajalal commented 6 months ago

Hello,

How can I create AR_data for the json file for custom synthetic dataset so that I can use it for CenterPose evaluator?

Currently, CenterPose evaluator expects AR_data as a field in the json annotation.

mujiwob commented 2 months ago

@monajalal Hi, do you know how to get AR_data now? I also encountered this problem when training in TAO centerpose.

Jason-xy commented 1 month ago

@monajalal @mujiwob Hi, did you find any solution to get AR_data for the json file? I have encountered the same issue while training in TAO centerpose.

mujiwob commented 1 month ago

@monajalal @mujiwob Hi, did you find any solution to get AR_data for the json file? I have encountered the same issue while training in TAO centerpose.

No. I found that AR_data is part of the Objectron dataset, but I don't know how to get it in the synthetic dataset by Isaac Sim. I just commented out the relevant code in centerpose_evaluator.py to skip this part, which made the training go on. However, the inference results after training are not good, I don't know if it is because of the lack of AR_data or my dataset is not good enough.

Jason-xy commented 1 month ago

I have bypassed this problem by assigning all fields in AR_data to 0, and then trained on my dataset of 20k images. finally, the results were good, so I guess AR_data has no impact on the current training process.

mujiwob commented 1 month ago

@Jason-xy So I think my poor results are probably due to the dataset. Is your dataset generated by Isaac Sim? If so, could you share the yaml setting files for generating the data and training? Thanks!