galidor/person-reid-autoencoder

Performing Person re-ID task on images being decompressed

Person-reID-Autoencoder

Performing Person re-ID task on images being decompressed with autoencoder. This is the soruce code for my master's thesis, written under supervision of Dr Krystian Mikolajczyk at Imperial College London, Faculty of Electrical and Electronic Engineering.

Table of Contents:

1. Data preparation
2. Training re-ID feature extractor

Data preparation

I performed all the experiments on CUHK03 dataset, availible here. You can simply download it from the website, unzip and put into a folder which I will refer to as data_path. You may also want to download new protocol for CUHK03, specifically file cuhk03_new_protocol_config_labeled.mat, available here. After these operations your data folder should look like this:

data_path

cuhk03_new_protocol_config_labeled.mat
cuhk03_release

cuhk03.mat
README.md

After prepairing the folder in such a manner, run the train_model_ResNet50.py program as follows:
python train_model_ResNet50.py --data_path /your/data/path --preprocess_dataset
Your data_path should then look like this:

data_path

cuhk03_new_protocol_config_labeled.mat
img
cuhk03_release

cuhk03.mat
README.md

All the images from CUHK03 "labeled" are stored in the folder img.

Training re-ID feature extractor

After finishing that you are ready for the next step which is training person re-ID feature extractor by running either train_model_ResNet50.py or train_model_PCB.py as follows (please note that certain arguments may be optional, just type python train_model_ResNet50.py --help for detailed description):

python train_model_ResNet50.py --data_path /your/data/path --batch_size 16 --model_path your/model/path --optim_step 20 --learining_rate 0.01 --epochs 50 --normalize --reranking  
python train_model_PCB.py --data_path /your/data/path --batch_size 16 --model_path your/model/path --optim_step 20 --learining_rate 0.01 --epochs 50 --normalize --reranking  

For me, training on GeForce GTX 1060 took around 1 hr per 40 epochs.
Your trained model will be saved in the selected directory, waiting for evaluation.

Autoencoder training