search

layumi / Pedestrian_Alignment

TCSVT2018 Pedestrian Alignment Network for Large-scale Person Re-identification
https://ieeexplore.ieee.org/document/8481710
MIT License
236 stars 66 forks source link

Maybe there is something difference between the paper and the code?? #14

Open zengqixun12 opened 7 years ago

zengqixun12 commented 7 years ago

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 commented 7 years ago

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 commented 6 years ago

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 commented 6 years ago

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 commented 6 years ago

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 commented 6 years ago

(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 commented 6 years ago

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.