CV-Assignment-2

Run an experiment

root/
  ├── .gitignore
  ├── utils.py
  ├── run_experiment.py
  ├── README.md
  ├── models/
  ├── data/
  ├── dataset/
  ├── preprocessing/
  └── experiments/
          └── experiment1/
              └── config.py

In order to create a new experiment, clone one of the experiments folder and edit the config.py file with your information. You must have a GPU for it to work otherwise it won’t run. Also, dowload this data https://www.kaggle.com/ashishpatel26/facial-expression-recognitionferchallenge and have the .csv in the data directory.

    python3 run_experiment.py -f <exp_folder_name>

Report

The report attached goes into much more details and is an interesting read(in our view 📚) ProjectReport.pdf

---

Architecture

Dataset

The original dataset is a grayscale image(originall in a different representation), as an experiment, we implement a model for facial landmarking to crop faces from images and compare accuracy. The motivation for cropping the faces is to reduce noise from the images and highlight useful features.

Results

The highest accuracy we received is ~80% on validation, which surpasses the current higest accuracy in the kaggle competition(from where the dataset is taken). Below are the results.

The Horror that lasts a lifetime

We decided to interpret the results from our CNN layers to understand the representations that each layer learns for an image.