Thanks for the authors' great work. I have set up the environment and reproduced similar results to the original paper. However, I find that in the second stage (fine-tuning), the author only uses the novel categories to fine-tune and evaluate on novel set. In TFA or FSCE etc. work, in the fine-tuning stage will use all the low-shot images of base categories and novel categories.
Thus, I don't know what if in the same paradigm that the results like. As your know, one of the core challenges in FSOD is to classify the confusingly base categories and novel categories, such like bird and ariplane.
I am now experimenting with a similar setup to TFA. I also hope the author can provide some explanation, I am really confused now. Thanks!
RuoyuChen10
commented
2 years ago
Thanks for the authors' great work. I have set up the environment and reproduced similar results to the original paper. However, I find that in the second stage (fine-tuning), the author only uses the novel categories to fine-tune and evaluate on novel set. In TFA or FSCE etc. work, in the fine-tuning stage will use all the low-shot images of base categories and novel categories.
Thus, I don't know what if in the same paradigm that the results like. As your know, one of the core challenges in FSOD is to classify the confusingly base categories and novel categories, such like
birdandariplane.I am now experimenting with a similar setup to TFA. I also hope the author can provide some explanation, I am really confused now. Thanks!
Best wishes!