Cross-View-Gait-Based-Human-Identification-with-Deep-CNNs
Cross-View-Gait-Based-Human-Identification-with-Deep-CNNs copied to clipboard
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
Introduction
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
Data
Prepare the dataset
get gait dataset from http://www.am.sanken.osaka-u.ac.jp/BiometricDB/GaitLP.html
,
use the Version 1, OULP-C1V1, http://www.am.sanken.osaka-u.ac.jp/BiometricDB/doc/OULP_Doc01a_SubsetStatistics_OULP-C1V1.pdf
,
download the dataset into path: ~/data/
, rename the directory to OULP_C1V1_Pack
,
you can see directories OULP-C1V1_NormalizedSilhouette(88x128)
and OULP-C1V1_SubjectIDList(FormatVersion1.0)
in ~/data/OULP_C1V1_Pack
.
Prepare training, validation and test dataset
protocol, described in paper: Cross-View Gait Recognition Using View-Dependent Discriminative Analysis.
In this experiments, training set described in the paper was split into traning and validation part, in OULP_setting/list_train.txt
and OULP_setting/list_val.txt
- change current directory to here
- move OULP_setting directory to
~/data/OULP_setting
, run command line:cp -r OULP_setting ~/data/OULP_setting
- run command line:
python preprocess_script/oulp_prepare.py ~/data/gait-oulp-c1v1
Running the code
Command line example to run the code
th main.lua -datapath ~/data/gait-oulp-c1v1 -mode train -modelname wuzifeng -gpu -gpudevice 1 -dropout 0.5 -learningrate 1e-3 -momentum 0.9 -calprecision 2 -calval 1 -batchsize 64 -iteration 2000000 >> main.lua.log
Result
you will see the validation average precision up to 92.50.
you will see results in main.lua.log which like below:
{
iteration : 2000000
seed : 1
loadmodel : ""
datapart : "test"
batchsize : 64
debug : false
gpu : true
gpudevice : 1
modelname : "wuzifeng"
calval : 1
momentum : 0.9
datapath : "/home/chenqiang/data/gait-oulp-c1v1"
gradclip : 5
dropout : 0.5
learningrate : 0.001
calprecision : 2
mode : "train"
}
2017-05-23 15:53:51[INFO] load data from /home/chenqiang/data/gait-oulp-c1v1/oulp_train_data.txt, /home/chenqiang/data/gait-oulp-c1v1/oulp_val_data.txt, /home/chenqiang/data/gait-oulp-c1v1/oulp_test_data.txt
2017-05-23 15:53:51[INFO] train data instances 06848, uniq 0856
2017-05-23 15:53:51[INFO] train data instances 00800, uniq 0100
2017-05-23 15:53:51[INFO] train data instances 07648, uniq 0956
nn.Sequential {
[input -> (1) -> (2) -> (3) -> output]
(1): nn.ParallelTable {
input
|`-> (1): nn.Sequential {
| [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
| (1): nn.SpatialConvolutionMM(1 -> 16, 7x7)
| (2): nn.ReLU
| (3): nn.SpatialCrossMapLRN
| (4): nn.SpatialMaxPooling(2x2, 2,2)
| (5): nn.SpatialConvolutionMM(16 -> 64, 7x7)
| (6): nn.ReLU
| (7): nn.SpatialCrossMapLRN
| (8): nn.SpatialMaxPooling(2x2, 2,2)
| }
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
(1): nn.SpatialConvolutionMM(1 -> 16, 7x7)
(2): nn.ReLU
(3): nn.SpatialCrossMapLRN
(4): nn.SpatialMaxPooling(2x2, 2,2)
(5): nn.SpatialConvolutionMM(16 -> 64, 7x7)
(6): nn.ReLU
(7): nn.SpatialCrossMapLRN
(8): nn.SpatialMaxPooling(2x2, 2,2)
}
... -> output
}
(2): nn.Sequential {
[input -> (1) -> (2) -> output]
(1): nn.CSubTable
(2): nn.Abs
}
(3): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> output]
(1): nn.SpatialConvolutionMM(64 -> 256, 7x7)
(2): nn.Reshape(1x112896)
(3): nn.Dropout(0.5, busy)
(4): nn.Linear(112896 -> 2)
(5): nn.LogSoftMax
}
}
2017-05-23 15:53:52[INFO] Number of parameters:1079906
2017-05-23 15:54:02[INFO] 00001th/2000000 Val Error 0.693167
2017-05-23 15:54:12[INFO] 00001th/2000000 Tes Error 0.693317
2017-05-23 15:54:21[INFO] 00001th/2000000 Tra Error 0.693182, 29
2017-05-23 15:54:31[INFO] 00065th/2000000 Val Error 0.693064
2017-05-23 15:54:40[INFO] 00065th/2000000 Tes Error 0.693260
2017-05-23 15:54:49[INFO] 00065th/2000000 Tra Error 0.693897, 27
Test
select the model file in trainedNet as a argument of -loadmodel when run main.lua, then you can see the test result in the redirected file. the best average recognition precision you can get is 88.29.
th main.lua -datapath ~/data/gait-oulp-c1v1 -mode evaluate -datapart test -modelname wuzifeng -gpu -gpudevice 1 -loadmodel ./trainedNets/wuzifeng_tra_0.6666_i7745.t7 >> main.lua.result.log