qingsong99 / person-reid-incremental

Incremental Learning in Person Re-Identification

Incremental Learning in Person Re-Identification

===========================================================

This repository contains code for our research. Paper can be found here, arXiv

Getting started

1. cd ~/PATH_NAME
2. Run git clone https://github.com/prajjwal1/person-reid-incremental
3. Install the specified dependencies, to install use pip3 install - requirements.txt
4. Follow the below mentioned steps for preparation of dataset and performing training

Prerequisites:

• OS: Linux/MacOS
• Pytorch>=0.3

Install Dependencies

• Installing Pytorch

Datasets

• Market 1501
• CUHK-03
• Duke MTMC

Dataset structure

This is the recommended file structure which was used

For preparation of Market1501

+-- Market1501
|   +-- bounding_box_test
|   +-- bounding_box_train
.......

For preparation of Duke MTMC

+-- dukemtmc-reid
  |   +-- DukeMTMC-reID
    |   +-- bounding_box_test
    |   +-- bounding_box_train
.............

For preparation of CUHK-03

+-- CUHK-03
|   +-- cuhk03_release
|   +-- splits_new_detected.json
|   +-- splits_new_labeled.json
.............

Covariance loss metric has been added to all the modules. You're required to change the flags as per phase as described in paper

Create a directory named as data/ and use the standard directory structure.

For training on Market1501:

$ python covariance_market1501.py

CUHK-03

$python covariance_cuhk-03.py

Be careful, Make sure that you are using the required split and flags, since training is CUHK-03 is more different than these two datasets. Settings specific to CUHK-03 have been marked with comments. The default mode loads detected images. Specify cuhk03-labeled if you wanna train and test on labeled images.

For training on Duke MTMC.

$ python covariance_duke.py

To use ensembling and training, use

$ python covariance_ensembling.py

In this case, you'll have to specify amongst which pipelines do you want to perform ensembling. If you get better results, please file a PR.

To perform training:

While executing make sure to correctly carry out training (Phase 1 and Phase 2) properly as mentioned to achieve incremental learning

When training, log file would be created in the /log directory.

Results:

No.	Dataset	Rank 1	Rank 20	maP
1	Market1501	89.3%	98.3%	71.8%
2	DukeMTMC	80.0%	93.7%	60.2%
3	Market1501	69.5%	92.8%	40.3%

No.	Dataset	Rank 1	Rank 20	maP
1	Market1501	89.3%	98.3%	71.8%
2	CUHK-03	-	-	-
3	Market1501	-	-	-

Takes around 8-9 hours to train the model for 950 epochs (convergence is usually achieved)

More benchmarks would be released soon.

Models

We used a ResNet50 along with different architecture of pipelines. We have used hybrid_convnet2.

To resume training

$ mkdir saved_models

Then specify this as per dir structure in the main module

SAVED_MODEL_PATH = 'saved_models/p1.pth.tar'
checkpoint = torch.load(SAVED_MODEL_PATH)
model.load_state_dict(checkpoint['state_dict'])

TO-DO

• Add support for Tensorboard (Pytorch)
• Usage of Random Erasing

Intel Technologies used

• Intel® Distribution of Python
• Intel® Math Kernel Library (Used internally by Pytorch)
• Intel® Nervana DevCloud (hyperparameter tuning,ablation studies,training model for quick experiments)

Pretrained weights

Will be made available soon.

Citation:

Please cite this, if you use our work

@ARTICLE{2018arXiv180806281B,
   author = {{Bhargava}, P.},
    title = "{Incremental Learning in Person Re-Identification}",
  journal = {ArXiv e-prints},
archivePrefix = "arXiv",
   eprint = {1808.06281},
 primaryClass = "cs.CV",
 keywords = {Computer Science - Computer Vision and Pattern Recognition},
     year = 2018,
    month = aug,
   adsurl = {http://adsabs.harvard.edu/abs/2018arXiv180806281B},
  adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}