I found the wh is hard-coded and fixed by default in gen_encoder_output_proposals(). Have you tried the settings of different size ratios? I noticed there is an option for learned wh but no results were provided in the paper (since the original DeformableDETR). Maybe it is not sensitive?
I found the wh is hard-coded and fixed by default in
gen_encoder_output_proposals()
. Have you tried the settings of different size ratios? I noticed there is an option for learned wh but no results were provided in the paper (since the original DeformableDETR). Maybe it is not sensitive?