Deep Bounding Box Regression using Keras

A Keras model to infer bounding box coordinates of synthetic black images with white boxes

Getting Started

• Clone the repo into your folder

  git clone https://github.com/sanjeev309/deep_bbox_regression_keras.git

• Launch jupyter notebook in the repo directory and open the notebook keras_bbox_regression.ipynb

Prerequisites

Keras, Matplotlib, OpenCV 3.2

Summary

Deep Models are good for tasks of classification. However they can also be utilised to solve regression problems.

Using Deep regression, a CNN model with Dense end nodes can be utilised to predict the 4 coordinates

of a bounding box namely (x1,y1) and (x2,y2)

For example:

Given the input image (100 * 100):

The ideal model predicts the bounding box coordinates :

[0.37, 0.31, 0.91, 0.81]

which when rescaled to the image corresponds to

[37, 31, 91, 81]

    i.e. (x1,y1) = [37,31] {Top-Left Coordinate}

         (x2,y2) = [91,81] {Bottom-Right Coordinate}

Author

• Sanjeev Tripathi - LinkedIn

License

This project is licensed under the MIT License - see the LICENSE file for details

Acknowledgements

• The database is generated by the tool available here : synthetic_bbox_regression_db_tool
• The loss function has been used from a different repo that I cannot locate. This Ack will be updated soon.