metadata

license: mit
task_categories:
  - object-detection

FTC Vision: 2024-2025 Dataset for Object Detection

This is the official dataset for FTC (FIRST Tech Challenge) object detection for the 2024-2025 season. Designed specifically for training object detection models using TensorFlow, this dataset is also provided in a RAW format for use in other frameworks or custom applications.

Dataset Overview

The dataset is structured to facilitate seamless integration into TensorFlow pipelines while maintaining flexibility for other use cases. It contains images and annotations in the following structure:

Annotations

Annotations are provided in XML VOC format, with bounding boxes drawn around each object of interest. Specifically, the dataset focuses on different-colored game pieces for the 2024-2025 season.
The annotations are further divided into train and val subsets to enable training and validation processes.
For TensorFlow users, the dataset includes:
- Train and Val TFRecord files for easy ingestion by TensorFlow pipelines.
- A label_map.pbtxt file to map class indices to human-readable class names (e.g., "red," "blue," "yellow").

Images

Images are split into train and val subsets, matching the corresponding annotations.

Each subset contains subdirectories named after their respective classes (e.g., red/, blue/, yellow/).

Example directory structure:

dataset/
  Images/
    train/
      red/
        img1.jpg
        img2.jpg
      blue/
        img3.jpg
      yellow/
        img4.jpg
    val/
      red/
        img5.jpg
      blue/
        img6.jpg
      yellow/
        img7.jpg

This structure ensures compatibility with TensorFlow's image_dataset_from_directory() and allows for quick model training.

Key Features

TensorFlow-Optimized Dataset: Includes TFRecord files and a label map for quick integration.
Raw XML VOC Annotations: Flexible format for users preferring other frameworks like PyTorch or custom data pipelines.
Class-Based Subdirectory Organization: Simplifies image classification and model evaluation workflows.
Bounding Boxes: Precisely annotated around game pieces of different colors, ensuring high-quality data for training object detection models.

Usage

TensorFlow:
- Use the provided TFRecord files (train.tfrecord and val.tfrecord) for model training.
- Load the dataset with TensorFlow's data pipelines and the provided label_map.pbtxt file.
Custom Frameworks:
- Use the raw XML annotations and class-organized image directories for model training in PyTorch or other frameworks.
Data Splits:
- The dataset is divided into training (80%) and validation (20%) sets to standardize evaluation.

Notes

Compatibility: The dataset is designed to be compatible with major machine learning frameworks.
Expandable: You can add additional game pieces, classes, or annotations to expand the dataset for future challenges.
Standardized Input Size: Images are resized to a consistent shape of (640, 640) for TensorFlow models but can be used at their original resolution in the RAW version.

This dataset is the ultimate resource for developing and training object detection models tailored to FTC's 2024-2025 game requirements. Let us know if you have any issues or questions!