hrzhou2 / AdaptConv-master

Adaptive Graph Convolution for Point Cloud Analysis
110 stars 12 forks source link

The different between xyz and xyz2_feat2 #10

Closed yeminglang closed 2 years ago

yeminglang commented 2 years ago

Dear authors, I notice that adaptive_feature = 'xyz' is selected for training, but according to you paper, it is adaptive_feature = 'xyz2_feat2' that should be used. By the way, it seems that adaptive_feature = 'asmy' never be used, because you chose if self.feat_mode == 'asym': rather than if 'asym' in self.feat_mode: in the code, and I think adaptive_feature = 'asmy' can't be use for its own. Maybe I do not get your idea... Can you please evaluate whether what I said is correct?

我注意到你们选择了 adaptive_feature = 'xyz' 进行训练,但根据你的论文,应该使用的是 adaptive_feature = 'xyz2_feat2'。还有个小问题,似乎从未使用过 adaptive_feature = 'asmy',因为您在代码中选择了 if self.feat_mode == 'asym': 而不是 if 'asym' in self.feat_mode:,因此我觉得 adaptive_feature = 'asmy' 不能单独使用(xyz和feat2就表示不了了)。也许我没很好地理解你们的想法......你能评估一下我说的是否正确吗?

Thanks.

Regards

hrzhou2 commented 2 years ago

Yes, 'asym' is the original asymmetric feature and is equal to 'xyz2_feat2'. You can use either of them. Also, these are architecture designs that are all suitable for building an adaptive conv layer. This is the reason why we set up such a feature type argument.

yeminglang commented 2 years ago

Thank you! Have a nice day!