I am implementing R-CNN in keras using this article https://towardsdatascience.com/step-by-step-r-cnn-implementation-from-scratch-in-python-e97101ccde55
In this article, The image is passed to the selective search and the proposed regions are generated
Then, using the annotation files, we test the IOU and get the object region only(region of interest) and also the background region
After that, trained with CNN.
In testing, the image is go through the selective search, and predict all the proposed regions, if the object present in any of the proposed regions, we can draw the bounding box.
It is simple, But, I experienced this situation - I tested an image which has exactly two objects and the result of the selective search has three purposed regions.
1st proposed region have 1st object
2nd proposed region have the 2nd object
3rd proposed region have both the two objects
I will draw the box around the object after predict the 1st and 2nd proposed region
In 3rd proposed region, the CNN predicts both the objects are present in the region.
If the CNN predicts the objects present in the 3rd proposed region then the box will be drawn around the two objects.
I do not want the box around two objects.
Like this image(It is a free hand drawn box around the object)
Is there any solution for this...
Could you help me to learn in this topic,
And, In most of the R-CNN articles, the authors mentioned to use regression for finding the bounding box. I am not fully understand this - Any suggestion about this regression
thanks