search

ethnhe / PVN3D

Code for "PVN3D: A Deep Point-wise 3D Keypoints Hough Voting Network for 6DoF Pose Estimation", CVPR 2020
MIT License
482 stars 105 forks source link

Bad results with my camera #103

Open RoyAmoyal opened 2 years ago

RoyAmoyal commented 2 years ago

hey,

I am using your pre-trained weights for the YCB dataset on my custom camera(Realsense) stream/pictures but I am getting bad results. I have changed the K intrinsics and the factor depth to my camera.

here you can see the best results I have managed to get for some objects but it's not accurate. image image

I am getting even worse performances when applying the net on other objects, sometimes the net localizes the objects in a total wrong locations and rotations.

How can I improve the results? maybe training on each individual YCB object will give me a better results? May it be because the weights for the whole YCB dataset are much smaller than the weights for each object in the Linemod Dataset?

Do you have any suggestions? Thanks!

RoyAmoyal commented 2 years ago

@ethnhe

and you maybe have any suggestions on how to create a new dataset for multiple objects together like the YCB dataset?