14  Midterm Project

This chapter describes a midterm project that concludes the CNNs for Scene Labeling part of the text. The project asks you to propose a research question relating to CNN-based scene classification, design a computational experiment to answer it, and report your findings as a research poster.

The project builds on the PatternNet workflow developed in Exercises 5.1, 10.1, and 11.1 and draws on the skills covered so far, including preparing DataSets and DataLoaders (Chapter 5), training and assessing models (Chapter 6 and Chapter 11), potentially improving models (Chapter 7), common CNN architectures (Chapter 12), and transfer learning (Chapter 13).

14.1 Midterm Project: CNN Scene Classification Experiment

14.1.1 Introduction

The goal of the midterm project is to design an experiment relating to CNN-based scene classification. You will build upon your results from Exercises 5.1, 10.1, and 11.1 by augmenting your model and/or training process and comparing the results to those obtained in Exercise 11.1. Some example modifications include:

  1. Using a different CNN architecture (e.g., ResNet, VGGNet, MobileNet, etc.)
  2. Reconfiguring your architecture to include residual connections
  3. Using different activation functions
  4. Augmenting the training process by using a learning rate scheduler
  5. Experimenting with different loss function configurations
  6. Implementing transfer learning (since the PatternNet images have red, green, and blue bands, ImageNet pre-trained weights can be used without modifying the first layer)
  7. Incorporating dropouts
  8. Incorporating data augmentations

You must propose a specific research question and augment your workflow to answer this question. Your instructor must approve your proposed research question.

14.1.2 Deliverables

  1. A Quarto (.qmd) file that contains all of the code from Exercises 5.1, 10.1, and 11.1, which will represent your baseline experiment, and the HTML webpage rendered from it.
  2. A Quarto (.qmd) file that contains all of the code for your new experiment/configuration, and the HTML webpage rendered from it.
  3. A research poster generated using graphics software, such as PowerPoint, Publisher, Inkscape, or Illustrator, and exported to a PDF file.

14.1.3 Grading

  1. Your research question is clearly stated, and your experiment is designed such that it addresses the research question.
  2. Your code is correct and commented throughout, and the experiment is correctly implemented.
  3. Research poster:
    1. The poster is well-designed and organized and uses space well. The poster should be presentation quality (ready to be taken to a meeting or conference).
    2. The input data and methods are defined/explained on the poster.
    3. The results are well presented.
    4. Key findings and conclusions are clear, well explained, and supported by your data.
    5. All included graphics/figures and tables are of high quality and well designed.

If you would like to pitch a midterm project using a different dataset, that is acceptable. However, you will still need to perform model comparisons, generate code, and produce a poster.