Person Re-identification using Siamese Neural Networks

This project tests out different models and datasets for person reidentification and evaluates its performance. Models used for this are -

1. ResNet50 Baseline as shown in Bag of Tricks Paper
2. OSNet
3. Custom MobilenetV3 Large backbone model

All the training data for this can be found on wandb on this link - https://wandb.ai/attention-boys/Mobilenet%20ReID%20LaST%20Script/workspace?nw=nwuserbhanushaliron This repository uses the training pipeline from https://github.com/shuxjweb/last Functions in the code are modified to enable wandb logging as well as new plotting functions and distance metrics are used. Additionally we have the train_model.py script, finetune.py, autorunner.sh and auto_evaluator.sh files added.

Environment Setup

Create a new pytorch environment using the given siamese_net.yml file. Run the following command

conda env create -f environment.yml
conda activate pytorch

Usage

To train the person ReID model, run the train_model.py script with appropriate command-line arguments specifying the training configuration. For example:

Training

python train_model.py --batch_size 64 --lr 0.00035 --model_name siamese --max_epochs 50 --train 1 --logs_dir /home/ronak/data/logs --dataset market1501 --log_wandb 1 --run_name siamese_market --data_dir /home/ronak/data/

Testing

python3 train_model.py --model baseline --train 0 --dataset market1501  --logs_dir /home/ronak/datasets/market1501/logs/baseline --data_dir /home/ronak/datasets/

Arguments passed

logs_dir

Path to store checkpoints.

run_name

Identifier for the run in Weights & Biases and also used for saving plots.

model

Select from "siamese" for a custom MobileNet-based network, "baseline" for ResNet50 backbone, and "osnet_x0_25" for OSNet with a 0.25 multiplier. Check the available models for other options.

dataset

Choose from "market1501", "dukemtmc", and "last". The code supports additional datasets not explored in this project; refer to the datasets __init__ file for more information on supported datasets.

data_dir

Path to the selected dataset.

Presentation Link - https://drive.google.com/drive/folders/1RfEx5Dv7uTjemIni9CiiSmViQzFH0fJi?usp=sharing

Pretrained Weights - https://drive.google.com/drive/folders/1Rmmy_538NdgABKeSNGC1O6ek79kmJIFl?usp=drive_link