How can I speed up the long inference time in first frame?

[shupinghu]

Dear authers, thanks for your contribution on model-free/based 6d pose estimation, this is actually an amazing work and really has a fantastic performance.

But I observed that the inference speed of first frame is slow, so I would want to speed up it. The inference time in first frame mainly costs on pose refiner and pose scorer, so I have tried to use less viewpoints or less iterations to do pose refining, but it would increase the pose error. Do you have any other ideas about it?

Since the inference time in first frame mainly costs on model inference, maybe a lighter model structure would make sense, could you please provide a lighter model or whole training code?

[wenbowen123]

https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_pose_estimation
this repo did some acceleration for that.