search

DecaYale / RNNPose

RNNPose: Recurrent 6-DoF Object Pose Refinement with Robust Correspondence Field Estimation and Pose Optimization, CVPR 2022
Apache License 2.0
160 stars 17 forks source link

Real time implementation #10

Closed Gan868611 closed 2 years ago

Gan868611 commented 2 years ago

hi i have two questions

  1. Is it true to say that i only need to provide init_pose for the first instance of the pose estimation, so given a stream of image, i only need to provide the init_pose for first image, is this understanding correct?
  2. how can i do real time implementation?
DecaYale commented 2 years ago
  1. Not very precise. For a single image, we only need one initial pose estimation to process. If you want to process a video, you could take the pose estimation from the previous frame t-1 to be the initial pose estimation of the frame t.
  2. There are still many aspects where you could further improve the speed. For instance, 1) you could re-implement the rendering parts with cuda like RePose; 2) you could implement the LM-optimization with cuda. Also, you could decrease the recurrent iterations and rendering iterations to achieve a tradeoff between the speed and accuracy. Actually, fewer refinement iterations are still enough for most application cases.