fucusy/cross-view-gait-based-human-identification-with-deep-cnns

Code for 2016 TPAMI(IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE) A Comprehensive Study on Cross-View Gait Based Human Identification with Deep CNNs

cross-view-gait-based-human-identification-with-deep-cnns's Issues

questions about the model you built

Hi, I am also from Beihang University. I am currently a junior student and also try to reproduce this paper.

However, I have some doubts about the model you built. It seems that the model you built is MT model. And the third convolution layer is devised to learn to perform 'subtraction' operation with outputs of two-channel convolution below.

However, in your code, it seems that you just subtract the outputs of the two-channel convolution manually and then do convolution on this subtraction.
local merge = nn.Sequential(); merge:add(nn.CSubTable()); merge:add(nn.Abs())
I suppose there is something wrong here. What do you think?

Overlarge output

Hi, I have successfully run the code without any error. However, in the main.lua.log, I got a lot of output like this:

2018-10-26 22:16:21[INFO] Number of parameters:1079906
2018-10-26 22:16:31[INFO] 00001th/2000000 Val Error 0.693183
2018-10-26 22:16:41[INFO] 00001th/2000000 Tes Error 0.693006
2018-10-26 22:16:52[INFO] 00001th/2000000 Tra Error 0.692732, 31
2018-10-26 22:17:02[INFO] 00065th/2000000 Val Error 0.693214
2018-10-26 22:17:12[INFO] 00065th/2000000 Tes Error 0.693444
2018-10-26 22:17:23[INFO] 00065th/2000000 Tra Error 0.693714, 30
2018-10-26 22:17:33[INFO] 00129th/2000000 Val Error 0.693021
2018-10-26 22:17:43[INFO] 00129th/2000000 Tes Error 0.693227
2018-10-26 22:17:54[INFO] 00129th/2000000 Tra Error 0.693305, 31
2018-10-26 22:18:56[INFO] forawrd 6412948-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49662798643112, same: true
2018-10-26 22:18:56[INFO] forawrd 6218860-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49696144461632, same: false
2018-10-26 22:18:56[INFO] forawrd 6118339-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49652144312859, same: false
2018-10-26 22:18:56[INFO] forawrd 6162019-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49622955918312, same: false
2018-10-26 22:18:56[INFO] forawrd 6057031-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49811816215515, same: false
2018-10-26 22:18:56[INFO] forawrd 6274860-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.4975294470787, same: false
2018-10-26 22:18:56[INFO] forawrd 6413431-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49766278266907, same: false
2018-10-26 22:18:56[INFO] forawrd 6325022-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49634358286858, same: false
2018-10-26 22:18:56[INFO] forawrd 6266656-IDList_OULP-C1V1-A-55_Gallery.csv, 6412948-IDList_OULP-C1V1-A-55_Probe.csv, dst, 0.49726143479347, same: false
......

And the file size has reached 4.3GB. I think this may not be normal. Could you give me some suggestions or tips? Thanks.

two more issues

many thx for the previous reply！

two more issues：

1. have you tested the model on the CASIA_B dataset？ can you achieve the 90% around accuracy？

2. have you observed the final performance difference between using nn.LogSoftMax and nn.SoftMax （since you used nn.LogSoftMax here）

average recognition precision

The average recognition precision 88.29 is for a 1 vs 1 verification problem or 1 vs N recognition problem?
Also, the around 90% accuracy on CASIA_B dataset (as illustrated in the paper) is for a 1 vs 1 verification problem or 1 vs N recognition problem?
BTW, can the author leave an email so that more details of the paper can be discussed with readers and researchers?