A issue about the paper

[JET-6LIU]

Hello author, I have some doubts about this paper. The paper mentions that general features are used to capture local class-defining features. Why use the CLS token to represent general features? Isn't the CLS token supposed to aggregate global information? Moreover, in ProtoPFormer, the CLS token is also used for global features.

[JET-6LIU]

The original text excerpt is as follows：

1. Generic features, through their localized focus on key class characteristics, provide clear discriminative signals for initial class separation. However, when differentiating between visually similar classes—such as distinguishing between dog breeds—specific features become more valuable, as they aggregate broader contextual patterns necessary for precise classification.

2. Generic features are extracted explicitly according to the backbone architecture:

• Transformer-based: the [CLS] token is used to represent generic features, following ProtoPFormer [39].
• CNN-based: Attention maps, interpreted as occurrence maps, are applied to extract generic features, inspired by XProtoNet [16]. Specific features, in contrast, are taken directly from the host architecture’s feature extraction pipeline without additional processing. without additional processing.

3. For ProtoPFormer, we adopted their approach of using the CLS token as a generic prototype and image tokens as specific prototypes.