Computer Vision Benchmarks

Two coursework projects for image classification and object detection in Advanced Topics in Computer Vision, University of Thessaly.

Overview

This repository contains two independent computer vision assignments:

Project	Folder	Description
Fashion-MNIST Classification	classification/	Compares shallow and deep dense neural networks on Fashion-MNIST.
Object Detection Benchmark	object_detection/	Compares YOLOv5m, Faster R-CNN, and SSD300 on 11 test images.

Each project includes its own source code, requirements file, README, and original report.

Preview

Classification training curves:

Faster R-CNN vs SSD300 comparison:

YOLOv5m detection output:

More saved detections are available in object_detection/docs/yolo_detections/.

Repository Structure

ComputerVisionBenchmarks/
|-- README.md
|-- .gitignore
|-- classification/
|   |-- README.md
|   |-- requirements.txt
|   |-- shallow_model.py
|   |-- deep_model.py
|   `-- docs/
|       |-- plots/
|       `-- classification_report.pdf
`-- object_detection/
    |-- README.md
    |-- requirements.txt
    |-- yolo_detect.py
    |-- faster_rcnn.py
    |-- ssd300.py
    |-- faster_rcnn_ssd_compare.py
    |-- images/
    |-- models/
    `-- docs/
        |-- model_comparisons/
        |-- yolo_detections/
        `-- object_detection_report.docx

Quick Start

Use a separate virtual environment for each project because the TensorFlow and PyTorch dependency stacks are large.

Classification

cd classification
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python shallow_model.py --no-plots

On macOS or Linux:

cd classification
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python shallow_model.py --no-plots

Object Detection

cd object_detection
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python yolo_detect.py --no-show

On macOS or Linux:

cd object_detection
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python yolo_detect.py --no-show

Results Summary

Classification

Metric	Shallow Model	Deep Model
Test accuracy	0.8846	0.8961
Validation accuracy	0.8923	0.8973
Training time	30.66 sec	64.47 sec

The deep model produced the best test accuracy, while the shallow model trained faster and stayed competitive.

Object Detection

Model	Main observation
YOLOv5m	Best balance of speed and detection quality.
Faster R-CNN	More exhaustive detections, but slower.
SSD300	Fast, but missed more objects and produced weaker labels.

Reports

• classification/docs/classification_report.pdf
• object_detection/docs/object_detection_report.docx

GitHub Notes

• Sample images and saved PNG result plots are included.
• Large model/checkpoint files such as *.pt, *.pth, and *.onnx are ignored.
• Generated outputs such as results/, runs/, and output/ are ignored.
• The local YOLO model file can be kept under object_detection/models/, but it should not be committed.

License

This project is licensed under the MIT License.

Author

Konstantinos Kiousis
Student ID: 2121129