Confusions about features used for similarity

[Holmes-GU]

Hi:

Thanks for sharing the codes. When reading file models/pointbert/point_encoder.py, I find it difficult to understand why the class token x[:, 0] is concatenated with the maximum of patches x[:, 1:].max(1)[0], as shown in the figure. Would you like to give some reasons for this operation?

Thanks.

[Tycho-Xue]

Hi @Holmes-GU , we followed how point bert is getting a single feature output for classification, I guess it's just a kind of way to aggregate information. other kind of aggregation should also work.