dataset_factory

This repository contains two groups of functions in C++

• data_factory_from_reid: generation of pairs and triplets from Re-Identification datasets with the division of the samples in train, validation and test sets, according to the protocol described by described in [3].

• data_factory_from_mot: generation of pairs, triplets and tracklets from Multi-object tracking datasets with the division of the samples in train, validation and test sets, according to the protocol described by described in [3].

The outputs are data txt files with labels, suitable for blobs creation to train deep networks with caffe.

Then, there is a set of scripts to create the blobs from the datasets using the generated data files (with the previous functions) - blobs_creation: generate caffe data blobs

Citation

Please cite dataset_factory_from_reid in your publications if it helps your research: @article{gomez2019balancing, title={Balancing people re-identification data for deep parts similarity learning}, author={G{'o}mez-Silva, Mar{'\i}a Jos{'e} and Armingol, Jos{'e} Mar{'\i}a and Escalera, Arturo de la}, journal={Journal of Imaging Science and Technology}, volume={63}, number={2}, pages={20401--1}, year={2019}, publisher={Society for Imaging Science and Technology} } Gómez-Silva, M. J., Armingol, J. M., & Escalera, A. D. L. (2019). Balancing people re-identification data for deep parts similarity learning. Journal of Imaging Science and Technology, 63(2), 20401-1.

Please cite dataset_factory_from_mot in your publications if it helps your research: @article{gomez2019transferring, title={Transferring learning from multi-person tracking to person re-identification}, author={G{'o}mez-Silva, Mar{'\i}a Jos{'e} and Izquierdo, Ebroul and Escalera, Arturo de la and Armingol, Jos{'e} Mar{'\i}a}, journal={Integrated Computer-Aided Engineering}, number={Preprint}, pages={1--16}, year={2019}, publisher={IOS Press} } Gómez-Silva, M. J., Izquierdo, E., Escalera, A. D. L., & Armingol, J. M. (2019). Transferring learning from multi-person tracking to person re-identification. Integrated Computer-Aided Engineering, (Preprint), 1-16.

Example of how to use data_factory_from_reid

This is an example of how to use data_factory_from_reid with PRID2011[1] and ViPER datasets[2].

string prid= "prid_dataset_directory"
get_samples(prid, 7,4);
train_val_test_division(prid, 100, 100, 100, 10, 100, 649, 100);
create_pair_data(prid, 100000, 10000, 1,4);
create_triplet_data_fixed_cam(prid, 50000, 5000);
create_triplet_data(prid, 50000, 5000);
create_test_data(prid);

string viper= "viper_dataset_directory"
get_samples(viper, 0, 3);
train_val_test_division(viper, 316, 316, 316, 10, 316, 316, 316);
create_pair_data(viper, 100000, 10000, 1,4);
create_triplet_data(viper, 50000, 5000);
create_triplet_data_fixed_cam(viper, 50000, 5000);
create_test_data(viper);

NOTE:be careful with PRID samples whose identification number is higher than 200, because different people in cam a and b are labbelled with the same number, from id 200. Alternative solution: remove samples with ID higher than 200 in cam_a set, they are not neccesarry in the training and test described in [3].

[1]Person Re-Identification by Descriptive and Discriminative Classification, Martin Hirzer, Csaba Beleznai, Peter M. Roth and Horst Bischof, In Proc. Scandinavian Conference on Image Analysis (SCIA), 2011

[2]D. Gray, and H. Tao, "Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features," in Proc. European Conference on Computer Vision (ECCV), 2008.

[3]Hirzer, M., Beleznai, C., Roth, P. M., and Bischof, H. (2011). Person re-identification by descriptive and discriminative classification. In Scandinavian conference on Image analysis, pages 91–102. Springer.

Example of how to use data_factory_from_mot

This is an example of how to use data_factory_from_mot with the MOT17 datasets from multi-object tracking challenge [4].

string mot= "mot_dataset_directory"
get_samples(mot, 0.7);
create_pair_data(mot, 500000, 50000, 16, 1, 0.9);
create_triplet_data(mot, 500000, 50000, 16, 1, 0.9);
create_tracklet_data(mot, 200000, 20000, 16, 16, 1, 0.9);

[4] https://motchallenge.net/data/MOT17/

ACKNOWLEDGMENTS

This work was supported by the Spanish Government through the CICYT projects (TRA2015-63708-R and TRA2016-78886-C3-1-R), and Ministerio de Educación, Cultura y Deporte para la Formación de Profesorado Universitario (FPU14/02143), and Comunidad de Madrid through SEGVAUTO-TRIES (S2013/MIT- 2713). We gratefully acknowledge the support of NVIDIA Corporation with the donation of the GPUs used for this research.