Skip to content

Latest commit

 

History

History
32 lines (23 loc) · 2.83 KB

dataset_creation.md

File metadata and controls

32 lines (23 loc) · 2.83 KB

Dataset Creation Strategy

Overview

To effectively train our few-shot object detection model, we devised a dataset creation strategy that combines real-world backgrounds with diverse objects. This approach ensures that our model learns to detect objects in a variety of contexts, enhancing its generalizability and robustness.

Steps in Dataset Creation

  1. Background Selection:

    • We began by selecting a series of background images from the web series Panchayat. These snapshots provide a realistic and varied set of scenes, ensuring that our model encounters a wide range of environmental contexts during training.

  2. Object Selection and Preparation:

    • We utilized objects from the Amazon Berkeley dataset, a rich resource containing various everyday items.
    • To integrate these objects seamlessly into our chosen backgrounds, we first removed the background from each object image, isolating the items of interest. This step involved using advanced image editing tools to ensure clean and precise object cutouts.

  3. Image Composition:

    • After preparing the background scenes and objects, we randomly placed 5-10 objects onto each background image. This random placement helps simulate different real-world scenarios and object arrangements.
    • Each composed image is thus unique, providing a wide array of training examples for the model.

  4. Augmentation:

    • To further enhance the diversity of our dataset, we applied various augmentation techniques. These include:
      • Rotation: Randomly rotating objects to different angles.
      • Scaling: Changing the size of objects to simulate distance and perspective.
      • Translation: Shifting objects within the image frame to different positions.
      • Color Jittering: Altering the color properties of objects and backgrounds to account for different lighting conditions.
    • Augmentation helps prevent overfitting and improves the model's ability to generalize to new scenes and objects.

  5. Fine-Tuning the Model:

    • With our augmented dataset ready, we fine-tuned a few-shot object detection model, specifically the DEtection Vision Transformer (DeVIT). DeVIT is well-suited for this task due to its powerful feature extraction and adaptability to new data with minimal examples.
    • Fine-tuning involved training the model on our custom dataset, adjusting it to accurately detect and classify the placed objects within the varied backgrounds.

By leveraging realistic scenes from Panchayat and diverse objects from the Amazon Berkeley dataset, combined with robust augmentation techniques, we created a comprehensive and diverse dataset. This meticulous dataset creation strategy enabled effective fine-tuning of the DeVIT few-shot object detection model, ensuring its capability to perform accurate object detection in a variety of real-world scenarios.