About logo classification task

[nTjing]

Hi Lin, you mentioned the logo classification task in your paper，I would like to ask if this classification task is to learn diverse feature vectors? One more question, how is this classification task implemented? Based on your paper, I think the classification model first capture the image features using ResNet and then connect the fully connected layer for classification. Is my understanding correct?

[lindsey98]

Hi, Yes, you are correct. During training, we train with normal softmax loss to make the representation embeddings to better distinguish intra-brand logos and inter-brand logos. During inference, we drop the classification head and only use the representation vector for comparison.

nTjing @.***> 于2023年10月18日周三 17:51写道：

Hi Lin, you mentioned the logo classification task in your paper，I would like to ask if this classification task is to learn diverse feature vectors? One more question, how is this classification task implemented? Based on your paper, I think the classification model first capture the image features using ResNet and then connect the fully connected layer for classification. Is my understanding correct? [image: image] https://user-images.githubusercontent.com/113650779/276200861-d4265bf8-a924-4e36-a377-bdc292f1e4c1.png

— Reply to this email directly, view it on GitHub https://github.com/lindsey98/Phishpedia/issues/20, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMJCOK3PF6AOKH6WFTKFDRTX76RAPAVCNFSM6AAAAAA6FHBAH2VHI2DSMVQWIX3LMV43ASLTON2WKOZRHE2DSMZTGE2DCOI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[nTjing]

In 5.5 section of the paper, Op2: How does a Siamese model trained with one-stage transfer learning and two-stage transfer learning perform logo comparison (see our discussion in Section 3.2)? What do one-stage transfer learning and two-stage transfer learning refer to respectively? The classification task is first trained on the Logo2K+ dataset, after which it is migrated to the list of target brands, and finally the fully connected layer is replaced with another fully connected layer with n output neurons, corresponding to the number of target brands fine-tuned. As a result, one-stage transfer learning is to migrate the learning from the categorization task to the target brand list. And two-stage transfer learning is to replace the full connectivity layer on top of the one-stage migratory learning for fine-tuning. Is this the right way to understand it?

[lindsey98]

One-stage is only trained on the Logo2K+ dataset. Two-stage is trained on Logo2K+ and then finetuned on our targetlist. Both are trained as a classification task, like what you mentioned: "the fully connected layer is replaced with another fully connected layer with n output neurons, corresponding to the number of brands".

nTjing @.***> 于2023年10月20日周五 16:23写道：

In 5.5 section of the paper, Op2: How does a Siamese model trained with one-stage transfer learning and two-stage transfer learning perform logo comparison (see our discussion in Section 3.2)? What do one-stage transfer learning and two-stage transfer learning refer to respectively? The classification task is first trained on the Logo2K+ dataset, after which it is migrated to the list of target brands, and finally the fully connected layer is replaced with another fully connected layer with n output neurons, corresponding to the number of target brands fine-tuned. As a result, one-stage transfer learning is to migrate the learning from the categorization task to the target brand list. And two-stage transfer learning is to replace the full connectivity layer on top of the one-stage migratory learning for fine-tuning. Is this the right way to understand it?

— Reply to this email directly, view it on GitHub https://github.com/lindsey98/Phishpedia/issues/20#issuecomment-1772296903, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMJCOK3TZ3T35MBYERH4BLTYAIYJDAVCNFSM6AAAAAA6FHBAH2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZSGI4TMOJQGM . You are receiving this because you commented.Message ID: @.***>

[Fujiaoji]

Hello Lin, for this I have a question. Since after finetuning based on Logo2k, you fine-tune again only for the last layer based on target list or all weights are changed.? Thanks

[lindsey98]

Hi Fujiao, All weights are changed.