Hi IDLab Team,
Thank you so much for providing your implementation on Github.
Could you clarify how you trained the original MTSD object detection models? I have a few questions about how you treated images during training and testing that I cannot determine from the provided code.
-
In section F.1, you describe a procedure that you applied to the dataset. From my understanding, this is the procedure that you used to add and manipulate the training images you used to perform the "Morphing" attack. Is this correct? If so, did you use the full MTSD dataset for training the non-morphing attack model?
-
For the non-morphing attack setup, how did you manage factors such as image size and classes? More specifically, given the size of the MTSD images, did you crop them to be compatible with the chosen model and then cut the testing data into segments? Moreover, given that the MTSD dataset has an "other-sign" class, did you treat these objects as background or include them as a class?
Thanks,
Kealan
Hi IDLab Team,
Thank you so much for providing your implementation on Github.
Could you clarify how you trained the original MTSD object detection models? I have a few questions about how you treated images during training and testing that I cannot determine from the provided code.
In section F.1, you describe a procedure that you applied to the dataset. From my understanding, this is the procedure that you used to add and manipulate the training images you used to perform the "Morphing" attack. Is this correct? If so, did you use the full MTSD dataset for training the non-morphing attack model?
For the non-morphing attack setup, how did you manage factors such as image size and classes? More specifically, given the size of the MTSD images, did you crop them to be compatible with the chosen model and then cut the testing data into segments? Moreover, given that the MTSD dataset has an "other-sign" class, did you treat these objects as background or include them as a class?
Thanks,
Kealan