I have a concern regarding the paper. It states that the Similarity-based Self-matching module is trainable and is optimized using the L1 loss. However, since both Fls and Flq are outputs of a frozen feature extractor, followed by masked pooling and averaging operations, I think it is more like a training-free module.
Thank you for your great work!
I have a concern regarding the paper. It states that the Similarity-based Self-matching module is trainable and is optimized using the L1 loss. However, since both Fls and Flq are outputs of a frozen feature extractor, followed by masked pooling and averaging operations, I think it is more like a training-free module.