During my first pass through the architecture for R-FCN, I believed that it was because their region proposal network generated regions of interest without performing the "small network slide" over every possible anchor (the same way that FRCNN performs.)

Now that I'm looking into implementations and attempting to replicate the results myself, I'm having trouble understanding how the R-FCN is faster than the FRCNN if their region proposal networks are the same.

Am I misunderstanding these architectures? I understand the idea behind k^2 feature maps but I'm not understanding how it speeds up the computation in practice.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.