{
  "fetched_at": "2026-10-19T17:29:04Z",
  "source_url": "https://www.kaggle.com/competitions/aml-2024-project-2/data",
  "pages": [
    {
      "name": "data-description",
      "content": "The dataset is based of the ``Traffic Sign Dataset'' with 54 different classes. For this work we will work with the dataset at the 75*75 pixel resolution and a subset of 9 classes that are the following: \n\n- 6 - Speed limit (70km/h)\n- 12 - Don't go left or right\n- 13 - Don't go right\n- 24 - Go right\n- 38 - Dangerous curve to the right\n- 39 - Dangerous curve to the left\n- 44 - Go left or straight\n- 46 - ZigZag Curve\n- 49 - Unknown5\n- 50 - Fences\n\nThe dataset set is composed of 277 images, with different distribution of examples per class. \n\n## Folders\n\n*   **TRAIN** - The training set\n*   **TEST** - the test set\n*   **sample_submission.csv** - a sample submission file in the correct format\n*   **labels.csv** - supplemental information about the data",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "Generative models are designed to learn the underlying distribution of a dataset and generate new data points that could plausibly come from the same distribution. Discriminative models make predictions based on specific traits, such as classifying photographs of animals. In contrast, generative models have the ability to create new instances of data, such generating a new image of an animal or a traffic sign.\n\nImage classification with deep neural networks tipically required a consi- derable amount of data to operate with good performance. When there is not that much data, data augmentation strategies are commonly used to tackle the issue. Moreover, aside from simple transformations, generative models have been also used to augment datasets in order to improve the performance of the base model.\n\nIn this work, you are tasked **to improve the performance of pre-determined image classifier by augmenting the dataset with images generated by gene- rative models**. The idea is to have the generative models to generate more samples for the training dataset with the overall intention to improve the performance of the classifier.\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Objective",
      "content": "The main objective is to analyse the dataset, prepare and train generative approaches that learn to generate more data for the training dataset of a predefined CNN. To do that you should attend to the following objectives:\n- Prepare the machine learning pipeline for the image classification da- taset;\n- Analyse the baseline CNN model provided;\n- Prepare the ML pipeline to train with augmented datasets, i.e. baseline dataset + generated samples;\n- Explore solutions to perform data augmentation via generative machine learning models to improve the performance of the current baseline model;\n- Explore solutions and compare the results using at least, the following Generative models:\n-- Any form of AutoEncoders;\n-- Any form of Generative Adversarial Networks (GAN);\n-- **(optional)** Any form of Diffusion models\n- Compare the results to the baseline model with and without the gene- rative samples added to training dataset.\n\n\nExploring other solutions than the listed ones that are suitable for the problem at hand and considered as extra work, can be as compensation points to cover problems in the listed ones above up to 10% compensation.\n\nExploring other solutions than the listed ones that are suitable for the problem at hand and considered as extra work, can be as compensation points to cover problems in the listed ones above.\n\n\n## Submission File\nFor each ID in the test set, you must predict a probability for the TARGET variable. The file should contain a header and have the following format:\n\n    ID,TARGET\n    2,0\n    5,0\n    6,0\n    etc.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "To evaluate the generalisation ability of the developed models, we are going to use a Kaggle competition. Note that it will not impact the final mark but rather will act as a way for you to access the progress you are making and evaluate the generalisation performance of your models.\n\n - Don't cheat!\n - Apply yourself!\n - Have fun!",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "abstract",
      "content": "We present here the second practical project, part of the students’ evalu- ation process of the Advanced Machine Learning course of the Master in Engineering and Data Science of the University of Coimbra. This work is to be done autonomously by a group of two students. The deadline for delivering the work is **2 June of 2024** via Inforestudante. The quality of your work will be judged as a function of the value of the technical work, the written description, and the public defence. All sources used to perform the work (including the code) must be clearly identified. The document may be written in Portuguese or in English, using a word processor of your choice1. The written report is limited to 8 pages long The document should be well structured, including a general introduction, a description of the problem, the approach, the experimental setup, an analysis of the results, and a conclusion. The report should follow the Springer LNCS format. The Latex and Word templates are available in the Support Mate- rial of the course. The final mark will be given to each member of the group individually. To do the work the student may consult any source he/she wants. Nevertheless, plagiarism will not be allowed and, if detected, it will imply failing the course. While doing the work and when submitting it, you should pay particular attention to the following aspects (whose relative importance depends on the type of work done):\n\n\n- description of the approach to the problem\n- description of the general architecture of the methods used;\n- description of the experiment, including a table with the parameters used which should allow full replication;\n- description of the evaluation metrics used for the validation: quality of the final result, efficacy, efficiency, diversity, or any other most appropriate;\n   \n\nDo not forget, besides what was just said, that it is fundamental: (1) to do a correct experimental analysis; (2) to do an informed discussion about the results obtained; (3) to put in evidence the advantages of the chosen alternative.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}