about code pointfusion
Opened this issue · 5 comments
dear wang ：
I want to do some comparative experiments of 6D pose estimation on PointFusion recently, but the code of PointFusion is not public yet. I see that there are some relevant experiments in your paper, so I wonder if you can send me the source code of PointFusion if possible. My email is yu.liang@siat.ac.cn
I will be very appreciated if you can help！
Best wishes.
Hi, the code of PointFusion belongs to the company that published the work and we also don't have access to it. But I do think you can modify DenseFusion to PointFusion. Basically, you can skip all the steps of pixel-level fusion and directly use the global feature from the pointnet as the input to the pose estimation network.
Hi, the code of PointFusion belongs to the company that published the work and we also don't have access to it. But I do think you can modify DenseFusion to PointFusion. Basically, you can skip all the steps of pixel-level fusion and directly use the global feature from the pointnet as the input to the pose estimation network.
Thanks for your reply. Actually, I have tried to modify DenseFusion to PointFusion. But I can't get as good a result as in your paper, and I don't know where the problem is. All I want is to calculate the add and AUC score of PointFusion on ycb-video, If you obtained these data in your previous experiments, I wonder if you can send me a copy. Thank you again for your reply！
I see. I really wish I can help but since it's too long ago, I could not find these results anymore. I would still recommend trying the modification idea from DenseFusion to PointFusion.
Thank you very much for your patience and recommendation. Actually，I followed the first part of PointFusion network and got the feature that size is nx3136, and then used the final regression part of DenseFusion to predict pose, but the result is not good enough. Do you have any suggestions?
Hi, I cannot find my modified PointFusion code at this moment. But I do remember I tried very hard to get the best PointFusion performance reported in the paper. It's really a long time ago to catch up all the changes. Did you use all the data augmentation steps for training the PointFusion?