search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.38k stars 648 forks source link

About ABCNet model size #229

Closed dagongji10 closed 3 years ago

dagongji10 commented 4 years ago

I have tried ABCNet use 2 pre-trained model ctw1500_attn_R_50.pth and tt_e2e_attn_R_50.pth, but I found there is some diffs between the size of these 2 pth file. ctw1500_attn_R_50.pth is 282M, but tt_e2e_attn_R_50.pth is 142M. So, when train ABCNet use CTW1500 and TotalText, the network is different? Why the model does not have same size?

Yuliang-Liu commented 3 years ago

Yes, they should be different.

CTW1500 use max_len: 100, which is because it is based on text-line level, i.e., one text instance could possibly contain many characters. The BezierAlign sampling grid is also different. For CTW1500, it is (8,128), while for the latter it is (8,32).

dagongji10 commented 3 years ago

@Yuliang-Liu thanks so much.