Maybe there is something difference between the paper and the code??

[zengqixun12]

First,thanks a lot for the align codes,it's quite important for me .

But ,after reading the paper twice and the code,I think there is something difference.

In the paper ,the Grid network are connected after the 4th ResBlock.However ,in the network restructed by the code,The grid network are connected after the 5th network......
Maybe there is something misunderstanding ？Thanks a lot for the answer. orz

[zengqixun12]

In mine mind, the ResBlock2-Resblock5 are the part of the stn network in order to figure out the θ.And the 'Affine network' in the paper,its work is to use the θ to calculate the target feature map and BP.

[layumi]

Hi @zengqixun12
The Grid network input is res4fx which size is 14x14x1024
Then I copy a 5th block as Grid network(net3) and add some conv layers.

[zengqixun12]

Thanks a lot for the answering!!
In the stn network,the input and the output should be the same size,so I think the Resblock2-Resblock5 has the the same function as the 'localisation-net' defined by the paper of the stn.The 'Grid network' and the 'Res3-Res4' in the PAN is corresponding to the 'localisation-net' in the stn.
Anyway,thank you very much for answering my questions carefully!
Waiting for your new paper!

[layumi]

Yes.
If you compare my code with the concept in original spatial transform network,
GridNet + (Res3+Res4 in base branch) serve as localisation net.

[TaihuLight]

(1) addpath CM_Curve in zzd_evaluation_res_faster.m is not need since that there are no dir of CM_Curve in your project?
(2) the name of function in train_id_net_res_market_align.m should be train_id_net_res_market_align, and not be train_id_net_vgg16?

[layumi]

Hi @TaihuLight
(1) This is the original evaluation api. For faster evaluation, I just removed it. So you can ignore it.
(2) The filename is more important. For experiment, I usually just copied the training code so the function name may be wrong but the filename is the key.