{
  "fetched_at": "2026-10-19T08:20:41Z",
  "source_url": "https://www.kaggle.com/competitions/data-centric-noisy-pascal-challenge/data",
  "pages": [
    {
      "name": "data-description",
      "content": "### Data\n\nThis competition uses a noisy version of the **PASCAL VOC 2007 dataset**, where the training data has been corrupted in phases to simulate real-world challenges. Your task is to improve data quality and object detection performance without any changes to the model.\n\n#### Files\n\n- **`voc2007/images/train/`**: Contains training images in `.jpg` format.\n- **`voc2007/labels/train/`**: Corresponding YOLOv5 format `.txt` files with annotations for each image.\n- **`voc2007/images/test/`**: Test images without labels.\n\n#### YOLOv5 Label Format\n\nEach `.txt` file contains:\n\n```\nclass_id x_center y_center width height\n```\n- **class_id**: Object class ID.\n- **x_center, y_center**: Bounding box center coordinates (normalized).\n- **width, height**: Bounding box size (normalized).\n\nExample:\n\n```\n6 0.485 0.533 0.122 0.250\n14 0.719 0.442 0.098 0.143\n```\n\n#### Prediction Task\n\nYou will predict bounding boxes and object classes for test images.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "### The Challenge\nIn real-world machine learning tasks, the quality of your data often matters more than the complexity of the model. This competition embraces the **data-centric AI** approach, where participants focus on improving the **data** rather than the model itself. Your objective is to enhance object detection performance on the widely-used **PASCAL VOC 2007 dataset** by correcting various types of data corruption.\n\nThe competition is structured around **four phases**, each designed to simulate real-world challenges faced when working with visual data. You’ll be working with the **YOLOv5 model**, and your goal is to **optimize its performance by enhancing, denoising, augmenting and correcting the dataset** across these phases.\n\n### Why Data-Centric AI?\nIn many computer vision tasks, tweaking model architectures can only get you so far. Often, noisy or low-quality data limits model performance, even with state-of-the-art algorithms. Data-centric AI puts the focus on improving **data quality** as a means to enhance model outcomes. By carefully managing contrast, noise, augmentation, and label accuracy, you’ll experience firsthand how **data quality is key to success** in object detection tasks.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "### Metric\n\nSubmissions will be evaluated using the **mean Average Precision (mAP50-95)** metric for object detection, a standard measure used to evaluate the performance of object detectors. The mAP50-95 score evaluates how well the predicted bounding boxes and class labels match the ground truth across various Intersection over Union (IoU) thresholds.\n\n### Submission File\n\nThe submission file for this competition is a CSV formatted as follows: Each row corresponds to a single image and contains four columns: `image_id`, `prediction_string`, `image_width`, `image_height`, `image_height` and 'Usage'. The `image_id` column specifies the unique identifier for each image (image file name), while `image_width` and `image_height` provide the dimensions of the image in pixels. The `prediction_string` column consolidates all detected bounding boxes for an image into a single, space-separated string. Each detection is represented by `{label confidence x_center y_center width height}`, where `label` indicates the class ID, `confidence` represents the detection confidence score, and `x_center`, `y_center`, `width`, `height` denote the bounding box normalized coordinates in YOLO format. The 'Usage' column is a fixed value set to \"Public\" because this competition includes only a public leaderboard. This column exists solely to meet Kaggle's requirements.\n\nFor example:\n\n```\n002084.jpg,7 0.9342 0.3527 0.5056 0.6568 0.9747 7 0.7538 0.4666 0.5264 0.9202 0.9442,500,375,Public\n007875.jpg,7 0.9515 0.4752 0.5047 0.9505 0.9732,414,500,Public\n\n```\n**The provided submission code is intended to prevent any potential conflicts, so you only need to prepare a trained YOLOv5 model.** Submit your predictions through the competition platform, and they will be automatically scored based on the ground truth annotations.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "## 1. General Competition Rules\n\n### Entry Requirements\nBy entering this competition, you agree to adhere to all the competition rules. You must create and submit your own work in compliance with the phases and requirements detailed below. Participants will be judged based on reproducibility and the effectiveness of their approach.\n\n### Team Size\n- The maximum team size is two (2) members.\n\n### Submission Limits\n- Participants are allowed to submit up to twenty (20) submissions per day.\n\n---\n\n## 2. Phase-Specific Rules\n\n### Phase 1 & Phase 2: Preprocessing Functions\n- Participants must create a preprocessing function to be applied to input data (train, validation, and test sets).\n- This function should address image enhancement and noise reduction, respectively, and **must** be applied before the data is fed into the model.\n- You are **not** allowed to change the model architecture, number of epochs, or any other hyperparameters at this stage.\n\n### Phase 3: Data Augmentation\n- In this phase, the focus is on optimizing data augmentation techniques.\n- You are allowed to modify augmentation-related hyperparameters to improve results.\n- However, any changes outside augmentation (e.g., model architecture) are prohibited.\n\n### Phase 4: Label Error Detection\n- Participants must implement an algorithmic approach to detect and correct label errors in the training data.\n- The approach must be **reproducible** and demonstrate improvements in model performance without manually altering the dataset.\n  \n### Important Considerations\n- Manual corrections to the data are not allowed. All improvements must come from reproducible methods applied programmatically.\n- The solutions must be robust, well-documented, and ensure that the steps taken are easy to replicate.\n\n---\n\n## 3. Reproducibility\n- Participants must submit clear documentation and code so that the preprocessing, augmentation, and label correction techniques can be replicated.\n  \nFailure to adhere to these rules may result in disqualification.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "abstract",
      "content": "Welcome to the **Data-Centric Object Detection Challenge**! In this competition, you'll be tasked with improving object detection performance by focusing on **data quality** rather than tweaking model architecture. Leveraging the Noisy PASCAL VOC 2007 dataset, your goal is to enhance the dataset in multiple phases, each focusing on a specific type of data corruption. \n\nThroughout the challenge, you will refine the dataset across four distinct phases:\n1. **Contrast Enhancement**: Address issues related to poor image contrast.\n2. **Noise Reduction**: Mitigate noise and artifacts using frequency domain techniques.\n3. **Data Augmentation**: Apply augmentation strategies to enhance model generalization.\n4. **Label Error Detection**: Correct errors in object localization and labeling.\n\nYour success in the competition depends on how well you can **enhance, denoise, augment and fix labels errors of the dataset** in these phases, ultimately improving the accuracy of object detection using the YOLOv5 model. Good luck, and let's see who can create the cleanest and most effective dataset! \n\n### Objective:\nImprove object detection performance by enhancing data quality through various methods and demonstrating your skills in data-centric AI.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}