yashkhare20 / imagebase

ImageBase: Automatic Image Annotation tool

An user-friendly framework designed to streamline the image annotation process. This system automatically annotates images while also supporting manual annotation. Additionally, it incorporates user-trained custom models to assist developers in projects requiring object detection and classification.

Run on local server:

1. Pull the code
2. Create a virtual environment with the `requirements.txt` file.
3. Run command "pip install -r requirements.txt"
4. Run command "python manage.py runserver"

-- Project Status: Completed

Project Intro & Objective

As image collections continue to grow in size, manual annotation becomes impractical, necessitating accurate and time-efficient methods. This project addresses this challenge by introducing an Automatic Image Annotation system designed to label images from a large pool of unlabeled images automatically.

Key Features:

1. Automated Annotation: Automatically labels images from unannotated datasets.
2. Image Processing: Utilizes image resizing for computational acceleration and employs image augmentation techniques.
3. Evaluation: Conducts experiments with various generic object detection algorithms and custom labeled datasets to assess system performance.
4. Future Development: Provides insights into further development directions for image annotation systems, drawing from both theoretical and experimental models.

How are companies labelling their data today?

• Organizations use a combination of software, processes, and people to clean, structure, or label data. In general, you have four options for your data labelling workforce:
• Employees - They are on your payroll, either full-time or part-time. Their job description may not include data labelling.
• Managed teams - They use vetted, trained, and actively managed data labellers (e.g. Cloud Factory).
• Contractors - They are temporary or freelance workers.
• Crowdsourcing - They use a third-party platform to access large numbers of workers at once.

Project Assumptions:

• Model is trained on objects/classes required by the user.
• Quality of image fulfils the threshold value for detecting objects in an image.
• The number of images for training crosses the threshold for successful prediction.

Project Limitations:

• The system will not work if low-quality images are given as input.
• The system will not predict the required class if the input number of that class do not cross the threshold.

Project Objectives:

• To provide a user-friendly graphical interface to input the image.
• User should be able to upload the image
• The system should detect objects present in the image.
• The system should be able to Annotate the image with the near accurate bounding box and their label.
• The system should export the annotations in required format.

Methods Used

• Machine Learning
• Computer vision

Technologies

• Python
• HTML/CSS
• Javascript
• OpenCV
• YOLO-V4 tiny
• Numpy, Pillow
• Django
• Ajax

Project Workflow - Automatic Annotation

Project Workflow - Manual Annotation

Model Architecture

Featured Notebook to Train Custom Model

• Yolov4-tiny