How can I train a model for localizing objects(classification not required) in Python

Question

I need to make a model that creates bounding box around objects(but does not classify them) for a competition. Which libraries or pre-trained models should I use. I need values of x1,x2(x1+w),y1,y2(y1+h), where all dimensions are measured from top left corner. The accuracy is measured by IOU(Intersection over Union).

Answer 1

The mainstream algorithms used today for finding the bounding boxes of objects in an image includes:

• R-CNN (and related Fast R-CNN and Faster R-CNN). This family of the algorithms actual does the localization (finding the bounding boxes) and classification in two separate steps. It finds the bounding box then does the classification. So if you don't need classification, just skip the second step.
• YOLO stands for 'You only look once'. The name comes from the fact this algorithm actually does the localisation and classification in one go.

Both algorithms can be implemented in the mainstream python deep learning library like tensor-flow or pytorch etc

Answer 2

Here is a good example how to do it in Keras using YOLO model.