Closed YueDai0514 closed 6 months ago
Full target point cloud - src point cloud = s2tflow.
Src point cloud + deformation = deformed src point cloud
Src point + Transformation + noise = Full target point cloud
Full target point cloud + crop = target point cloud
Thank you very much for your reply. I'd like to ask for some more advice on how to organize the directory of data, the rot, trans, s2t_flow, src_pcd, tgt_pcdfull, , f_mask, scale = self.entry2data(index) and entry2data to read the data in the entry, so does the entry data need to be ready beforehand, and is the data of these How is the format organized, please excuse my confusion.
For me, I prefer to process them first and then load during training and testing
Thanks for the reply, I'll look into it again
Hi, In the demo implementation, RE = tgt_full - tgt_full_pred, so how is the ground truth error obtained? Looking forward to your reply.
tgt_full from ground truth. tgt_full_pred = R_pred.src + t_pred
Also, applying point cloud matching in this field is still an open problem. The current evaluation metrics are designed for the simulation cases. In my humble opinion, we need better evaluation metrics and a more realistic dataset, which I have been working on. So, I don't think you need to follow my routine. You can choose a way that makes sense for you. The most important message from this project is that this approach is very promising and deserves more attention. Best luck!
This is very inspiring to me, thank you very much for your valuable advice.
Hi, I would like to know how the ["s2t_flow"] displacement field in the code is obtained, is it also generated by V2S-Net or some other other processing.