Vehicle Re-Identification 🚗

Hello! This is a coursework repository for EEEM071: Advanced Topics in Computer Vision and Deep Learning. This is a fork of the original course work repository with some QoL improvements:

• Weights & Biases Integration 📊 for seamless experiment tracking
• Makefile support 🛠️ for running the experiments
• Poetry 📜 for dependency management

If you are a University of Surrey student, you are welcome to use this project as a learning resource and reference for your coursework. A simple credit to the OC (wee! that's me, Frank) would be greatly appreciated. However, please note that submitting this work as your own academic assignment is not permitted and may lead to academic misconduct penalties. Just make sure you're submitting your orignal work.

Code Overview 🧑‍💻

flowchart LR
    A[Data Loading] --> B[Model Setup]
    B --> C[Training]
    C --> D[Evaluation]
    
    subgraph A[Data Loading]
        A1[src.datasets.veri] -- data augmentation --> A2[src.transforms]
        A2 -- create train, test dataloaders --> A3["src.data_manager.return_dataloaders()"]
    end
    
    subgraph B[Model Setup]
        B1["src.models.init_model()"] -- setup xent and htri losses --> B2[src.losses]
        B2 -- configure optimizer and scheduler --> B3[src.optimizers + src.lr_schedulers]
    end
    
    subgraph C[Training]
        C1["optimizer.step()"] -- calculate losses --> C2[xent_loss + htri_losss]
        C2 -- backpropagation --> C3["loss.backward()"]
        C3 -- forward pass --> C1
    end
    
    subgraph D[Evaluation]
        D1["qf, q_pids, q_cam_ids, gf, g_pids, g_cam_ids"] -- calculate distance matrix --> D2[distmat]
        D2 -- Compute cmc and mAP --> D3["evaluate(distmat, q_pids, g_pids, q_camids, g_camids)"]
    end

Loading

Directory Structure 🌳

VeRi
├── Makefile
├── README.md
├── args.py
├── assets
├── docs
├── experiments
│   ├── extra
│   ├── section1
│   ├── section2
│   └── section3
├── main.py
├── poetry.lock
├── pyproject.toml
├── src
│   ├── data_manager.py
│   ├── dataset_loader.py
│   ├── datasets
│   ├── eval_metrics.py
│   ├── losses
│   ├── lr_schedulers.py
│   ├── models
│   ├── optimizers.py
│   ├── samplers.py
│   ├── transforms.py
│   └── utils
└── train.sh

All relevant papers are stored in the docs directory. A sample training script is provided in train.sh. I'm also using sourcetrail to visualize the code structure. It's been discontinued but still works fine. You can download it here.

Running the Experiments 🏃

You can find the experiments for different sections under the experiments folder. Please run it using the Makefile in the root directory.

make section1

These experiments are run on my own 4080 Super FE, with the resutls both stored locally and on W&B.

Experiment Results 📈

All experiment results are stored on this W&B workspace. The summary of the best configurations for each section is as follows:

Section	Hyperparameter	Best Configuration	mAP (%)	Rank-1 (%)
Section 1	Network architecture	resnet50_fc512	54.54	87.01
Section 2	Data augmentation	Default + Random Perspective	54.61	87.07
Section 3.1	Learning rate	0.0001	64.96	91.06
Section 3.2	Batch size	64	65.85	91.95
Section 3.3	Optimizer	AMSGrad	65.85	91.95

You can also find detailed results on my W&B report here.

Checklist 📝

• Section 1: Different network architectures
• Section 2: Data augmentation
• Section 3: Hypyerparameter tuning
• Section 4: Summary

License 📃

This project is licensed under the MIT License. This means:

• Attribution: You must give appropriate credit to the original author (me, Frank Lü) if you use or modify this project.
• Freedom: You are free to use, modify, and distribute this project, including for commercial purposes, without restriction.
• No Warranty: This project is provided “as-is” without any warranty or liability on the part of the author. You can read the full license in the LICENSE file.