search

wenbowen123 / BundleTrack

[IROS 2021] BundleTrack: 6D Pose Tracking for Novel Objects without Instance or Category-Level 3D Models
Other
612 stars 66 forks source link

keypoints matches are not sampled? #64

Closed royzhang12 closed 1 year ago

royzhang12 commented 1 year ago

Hi @wenbowen123,

thank you for your excellent work. Just have a small question about the implementation. I found that in the code when running the RANSAC between two images, the registration sample of 3 pairs of keyspoints seems not used, while keypoint matches created by findCorresbyNN are directly used as input of RANSAC.

Specifically, in the runRansacBetween(frameA, frameB) method of the SiftManager class, the sample_to_try(which refers to a list of samples consisting of 3 pairs of matches) seems not to be used. (I actually try to delete this part before runRansacMultiPairGPU({{frameA, frameB}}) in the method and still get the same result).

Is there any part I ignored or actually directly do the RANASAC on keypoints matched without sampling them in a group of samples consisting of 3 pairs has a better effect in inference?

royzhang12 commented 1 year ago

I found out that the sample process is actually achieved in the cuda_ransac.cu.