Vehicle Re-Identification with Deep Learning

Systematic hyperparameter tuning and architecture benchmarking for vehicle re-identification — evaluated on the VeRi-776 dataset.

This project was completed as part of an MSc Artificial Intelligence programme at the University of Surrey (Module: Advanced Topics in Computer Vision and Deep Learning).

Note: The training codebase was provided by the University of Surrey and is subject to copyright — it is not included in this repository. This repository documents the experimental methodology, hyperparameter configurations, and results.

---

Task Overview

The goal was to re-identify vehicles across multiple cameras using deep learning models trained with a combined cross-entropy and hard triplet loss. The assignment was structured as a sequential hyperparameter tuning exercise across four sections.

Dataset: VeRi-776 — 576 vehicle identities, 37,778 training images, 1,678 query images, 11,579 gallery images captured across 20 cameras. Performance is measured using mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) Rank-1/5.

Loss functions: Cross-entropy loss + hard triplet loss (AMSGrad optimiser, unless otherwise noted).

---

Section 1 — Architecture Comparison

Three architectures were trained with identical default hyperparameters (AMSGrad, lr=3×10⁻⁴, batch size=64, 10 epochs) and evaluated on the test set.

Experiment	Model	mAP (%)	Rank-1 (%)	Rank-5 (%)	Params (M)	Feature Dim	Training Time
Q1 (default)	MobileNetV3-Small	44.5	80.2	90.9	0.93	576	34 min
Q2	ResNet50	51.1	83.4	92.4	23.5	2048	2 h 9 min
Q3	VGG16	18.2	53.2	67.6	14.7	25088	2 h 56 min

Key finding: ResNet50 outperformed MobileNetV3-Small by +6.6% mAP (+15%). Its residual connections improve gradient flow and its 2048-dimensional feature map captures richer vehicle details. VGG16 performed worst despite its large feature vector — the 25,088-dimensional representation likely caused overfitting, and the absence of skip connections hinders gradient flow. The lightweight MobileNetV3-Small offered a strong accuracy/speed trade-off at 34-minute training.

---

Section 2 — Data Augmentation Experiments

Starting from the default augmentation (random horizontal flip + Random2DTranslation), additional techniques were appended one at a time, then combined.

Configuration	mAP (%)	Rank-1 (%)	Rank-5 (%)
Default	44.5	80.2	90.9
Default + Colour Jitter	44.6	81.0	90.8
Default + Random Erase	46.3	80.0	90.7
Default + Random Erase + Colour Jitter (best)	46.6	81.0	91.0

(MobileNetV3-Small, AMSGrad, lr=3×10⁻⁴, batch size=64, 10 epochs)

Key finding: Colour jitter alone produced a negligible gain (+0.1% mAP), indicating that moderate colour variation does not provide meaningful additional signal. Random erasing delivered a clear +1.8% improvement by masking regions of the vehicle, forcing the model to focus on other discriminative parts. Combining both techniques yielded the highest mAP of 46.6%, as colour diversity strengthens the features learned over the unmasked patches.

---

Section 3 — Hyperparameter Exploration

3.1 Learning Rate

Five learning rates were evaluated. All other settings were held at default.

Learning Rate	mAP (%)	Rank-1 (%)
1×10⁻⁵	21.2	46.5
5×10⁻⁵	36.0	64.9
1×10⁻⁴	42.2	75.3
3×10⁻⁴ (default — best)	46.3	80.0
1×10⁻³	41.0	79.0

(MobileNetV3-Small, AMSGrad, batch size=64, 10 epochs)

Key finding: lr=3×10⁻⁴ was optimal. Very low learning rates (≤5×10⁻⁵) failed to converge within 10 epochs. A high learning rate (1×10⁻³) converged quickly but overshot stable minima, generalising poorly. The default rate provided the best balance between convergence speed and stability.

3.2 Batch Size

Using the best learning rate (lr=3×10⁻⁴) and best augmentation, five batch sizes were evaluated.

Batch Size	mAP (%)	Rank-1 (%)
32	47.0	82.7
48 (best)	47.6	82.4
64 (default)	46.3	80.0
72	47.0	81.8
84	47.2	80.8

(MobileNetV3-Small, AMSGrad, lr=3×10⁻⁴, 10 epochs)

Key finding: Batch size 48 achieved the highest mAP of 47.6%. Smaller batches introduce gradient noise that acts as an implicit regulariser, improving generalisation. Very small batches (32) and very large batches (≥64) both performed slightly worse — the former can be unstable, while the latter reduces gradient noise and sharpens minima.

3.3 Optimiser

Fixing the best learning rate (lr=3×10⁻⁴) and batch size (48), SGD was compared against the default AMSGrad.

Optimiser	mAP (%)	Rank-1 (%)
AMSGrad (default)	47.6	82.4
SGD	19.9	44.4

(MobileNetV3-Small, lr=3×10⁻⁴, batch size=48, 10 epochs)

Key finding: Switching to SGD caused a severe drop to 19.9% mAP. Without adaptive moment estimates, SGD with momentum failed to navigate the loss landscape effectively within 10 epochs. AMSGrad's per-parameter adaptive learning rates are clearly better suited to this task.

---

Summary of Findings

Section	Factor	Best Configuration	mAP (%)	Rank-1 (%)
1	Architecture	ResNet50	51.1	83.4
2	Data Augmentation	Default + Random Erase + Colour Jitter	46.6	81.0
3.1	Learning Rate	3×10⁻⁴	46.3	80.0
3.2	Batch Size	48	47.6	82.4
3.3	Optimiser	AMSGrad	47.6	82.4

The best single result across all experiments was ResNet50 with default hyperparameters (mAP = 51.1%). For MobileNetV3-Small, combining optimal augmentation, learning rate, and batch size progressively raised mAP from 44.5% to 47.6%.

---

Experiment Configuration

All experiments were run using the university-provided training script (main.py). The baseline configuration is shown below; individual arguments were modified for each subsequent experiment as described above.

python main.py \
  -s veri \
  -t veri \
  -a mobilenet_v3_small \
  --root path/to/VeRi \
  --height 224 \
  --width 224 \
  --optim amsgrad \
  --lr 0.0003 \
  --max-epoch 10 \
  --stepsize 20 40 \
  --train-batch-size 64 \
  --test-batch-size 100 \
  --save-dir logs/mobilenet_v3_small-veri

See train.sh for all experiment configurations with results annotated as comments.

---

Tools & Environment

Tool	Purpose
Python / PyTorch	Model training and evaluation
MobileNetV3-Small	Default lightweight CNN backbone
ResNet50	Best-performing architecture (residual CNN)
VGG16	Plain CNN comparison
VeRi-776	Vehicle re-identification dataset (20 cameras, 576 IDs)
Google Colab / University HPC	Training compute platform (GPU)

---

References

• He, K., et al. (2016). Deep residual learning for image recognition. CVPR 2016.
• Howard, A., et al. (2019). Searching for MobileNetV3. ICCV 2019.
• Keskar, N. S., et al. (2017). On large-batch training for deep learning: Generalisation gap and sharp minima. ICLR 2017.
• Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimisation. ICLR 2015.
• Liu, X., et al. (2016). Large-scale vehicle re-identification in urban surveillance videos. ICME 2016.
• Schroff, F., et al. (2015). FaceNet: A unified embedding for face recognition and clustering. CVPR 2015.
• Simonyan, K., & Zisserman, A. (2015). Very deep convolutional networks for large-scale image recognition. ICLR 2015.

---

MSc Artificial Intelligence, University of Surrey