Repository navigation
PostProcessing Spatial Integration #158
|
Good day, Happy new year! I have been wondering the reasoning that postprocessing spatial integration, which should drastically increase detection accuracy, is done outside the architecture (training/validation), but not done as part of the final layers before feeding into loss? Thank you, |
Answered by
Haydnspass
Jan 4, 2022
Replies: 1 comment
|
Hey @tsuijenk, I think that's a standard procedure in many DL algorithms. oftentimes post-processing can not be backpropped through and it the other point is that training should ensure a good prediction that needs as little post-processing as possible. If you had a backproppable post-processing algorithm you can often end up in situations in which the model predictions worsens because the model "knows" that the predictions will be refined. |
0 replies
Answer selected by
tsuijenk
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hey @tsuijenk,
I think that's a standard procedure in many DL algorithms. oftentimes post-processing can not be backpropped through and it the other point is that training should ensure a good prediction that needs as little post-processing as possible. If you had a backproppable post-processing algorithm you can often end up in situations in which the model predictions worsens because the model "knows" that the predictions will be refined.