Deep neural network model introducing new novel matching layer called 'Normalized correlation' layer. This repository contain information about the datasets used, implementation code. The paper titled "Deep Neural Networks with Inexact Matching for Person Re-Identification" is accepted in NIPS-2016. You can find the paper here
A short video that was prepared for NIPS-2016:
Poster of the paper can be found here.
An independent implementation in Keras by sumedhpendurkar@github can be found here & an explanation about the implementation is here.
| Dataset name | Description |
|---|---|
| CUHK03 (Labeled & Detected) | A collection of 13,164 images of 1360 people captured from 6 different surveillance cameras, with each person observed by 2 cameras with disjoint views. The dataset comes with manual ("Labeled" dataset) and algorithmically ("Detected" dataset) labeled pedestrian bounding boxes. The dataset can be obtained from here |
| CUHK01 | A mid-sized dataset with 9884 images of 971 identities (4 images per identitiy). The dataset can be obtained from here |
| QMULGRID | A small sized dataset with 250 identities observed from 2 cameras (2 images per identity). There are 775 unmatched identities (1 image per identity) to be included as part of Gallery images during test run. Dataset can be downloaded from here |
More dataset links can be found here
The code development has been done in the environment as mentioned below:
Ubuntu 14.04 LTS
- Torch (with packages (nn, cunn, cutorch, image) installed by default, as well as some special packages such as matio)
- Matlab (R2014a version) - for some data preprocessing scripts
NVIDIA-SMI 352.41 Driver Version: 352.41 with GeForce GTX TITAN GPUs
The data-preprocessing scripts are written mainly using Lua (torch), Matlab.
cd to the folder datasets/cuhk03
Follow the steps below:
- Download the CUHK03 dataset after filling out a google form available in Rui Zhao's Homepage
- Place the file (cuhk-03.mat) in the folder ./datasets/cuhk03
- Execute the script "execute-all.sh"
If everything goes well, there will be 4 folders ("labeled", "labeled_testsets", "detected", "detected_testsets") available in "datasets/cuhk03" folder.
cd to the folder datasets/cuhk01_test100
Follow the steps below:
- Refer to this page to get the CAMPUS person re-id dataset
- Extract the zip file (CAMPUS.zip) in the folder ./datasets/cuhk01_test100 (now, this folder should contain a folder named 'campus')
- Execute the script "execute-all.sh"
If everything goes well, there will be 2 folders ("train", "test") available in "datasets/cuhk01_test100" folder.
cd to the folder datasets/cuhk01_test486
Follow the steps below:
- Refer to this page to get the CAMPUS person re-id dataset
- Extract the zip file (CAMPUS.zip) in the folder ./datasets/cuhk01_test486 (now, this folder should contain a folder named 'campus')
- Execute the script "execute-all.sh"
If everything goes well, there will be 2 folders ("train", "test") available in "datasets/cuhk01_test486" folder.
cd to the folder datasets/qmulgrid
Follow the steps below:
- Download the QMULGRID dataset from here
- Extract the zip file (underground_reid.zip) in the folder ./datasets/qmulgrid (now, this folder should contain a folder named 'underground_reid')
- Execute the script "execute-all.sh"
If everything goes well, there will be 3 folders ("train", "test", "additionalgallery") available in "datasets/qmulgrid" folder.
The options used during training are available in the file "src/opts.lua"
opt.dataset = 'cuhk03'; -- cuhk03 | others
opt.dataType = 'detected' -- labeled | detected | ''
--------------
opt.datasetname = 'cuhk03' -- cuhk03 | cuhk01_test100 | cuhk01_test486 | qmulgridThe options opt.dataset and opt.dataType should be changed carefully according to the dataset.
opt.dataset = 'cuhk03'; -- cuhk03
opt.dataType = 'detected' -- labeled | detected
--------------
opt.datasetname = 'cuhk03'opt.dataset = 'others';
opt.dataType = ''
--------------
opt.datasetname = 'cuhk01_test100' -- cuhk01_test100 | cuhk01_test486 | qmulgridopt.modelType = 'normxcorr' -- normxcorr | cin+normxcorropt.GPU = 1 -- default GPU to hold the original copy of model
opt.nGPUs = 3 -- the total number of GPUs to be used during training The other options should be left as it is (Reason: they are for future use / error handling is not proper for them).
Use the file 'doallTest.lua' for testing the model for any particular trained model.
All you have to do is, to change the options in the file 'doallTest.lua'. Note that, 'doallTest.lua' file does not depend on 'opts.lua'.
opt.dataset = 'cuhk03'; -- cuhk03 | others
opt.datasetname = 'cuhk03' -- cuhk03 | cuhk01_test100 | cuhk01_test486 | qmulgrid
opt.datapath = '../datasets/' .. opt.datasetname .. '/'
opt.dataType = 'detected' The meaning of these options is as same as the training options mentioned above. As per these options, the data will be read from appropriate dataset directory, tests will be carried out for 10 times and average CMC (Cumulative Matching Characteristics) will be calculated at the end of the tests.
MODEL_PATH = '<absolute/relative path>' -- exact path of the modelThe test log will be created in the same path of the model.
The other options should be left as it is (Reason: they are for future use / error handling is not proper for them).
To execute a particular file (doall.lua,doallTest.lua), first cd to the folder ./src folder.
After making sure that the preprocessing of the data is done correctly and setting all the options accordingly, execute the file doall.lua as,
th doall.luaThe trained model(s) will be stored (after every epoch) in the folder './scratch'.
After setting the options for test in 'doallTest.lua' file, execute it as,
th doallTest.lua| File path | purpose |
|---|---|
| src/doall.lua | the main file to be executed for training |
| src/doallTest.lua | the test execution file to get CMC (Cumulative Matching Characteristics) percentages for a particular model and dataset) |
| src/model_<xxxx>.lua | defines a particular type of model |
| src/trainMultiGPU.lua | training subroutine using Multi GPUs |
| src/dataForTests.lua | data reading procedures for 'test'; reads the data depending on the options set in doallTest.lua |
| src/data.lua | data reading procedures for training and validation; reads the data depending on the options set in opts.lua |
| src/log.lua | logging helper functions. copied from here |
| src/rankScores.lua | to rank the gallery images based on Softmax scores |
| src/doallTest.lua | code to carry out test for a particular trained model and dataset |
| src/doall.lua | code to carry out training for a particular dataset and type of model sepcified in opts.lua |
| src/loss.lua | defines loss/objective/criterion function (negative log likelihood - NLL) used during training |
| src/opts.lua | contains options for training |
| src/test.lua | testing subroutine to test for probe images and rank the gallery based on softmax classifier scores (applicable for both validation and testing) |
| src/utilities.lua | contains miscellaneous multi-purpose utility functions that are used in other code files |
| src/modules/* | contains code files for Normalized correlation Matching layer (NormCrossMapCorrelation.lua), Ahmed at al.'s Cross Input Neighborhood layer (CrossInputNeighborhood.lua), Parallel Multi GPU model training package (DataParallelTableForSiamese.lua) inspired and adapted from Facebook's torch [DataParallelTable.lua] (https://github.com/torch/cunn/blob/master/DataParallelTable.lua) and CUDA C++ implementation of modules |
@inproceedings{subramaniam2016deep,
title={Deep Neural Networks with Inexact Matching for Person Re-Identification},
author={Subramaniam, Arulkumar and Chatterjee, Moitreya and Mittal, Anurag},
booktitle={Advances In Neural Information Processing Systems},
pages={2667--2675},
year={2016}
}