Some questions about memory consumption and running time in PETR

[Asafgendler]

Hey, thanks for the great paper,

I had two issues i wanted to discuss about the PETR model.

1. Since the joint decoder runs simultaneously on all of the pose queries, it's effectively have a batch size of num_pose_queris * batch_size. Doesn't it cause a huge memory consumption problem?

2. Following the previous question, doesn't activating the joint decoder on so many queries (num_pose_queris * batch_size * num_key_points) causes a very large latency?

Thanks,