Detecting object borders PRECISELY

Question

I need to detect object (coin) in a photo, which is to draw a rectangle around it. I tried Mark-RCNN, Retina, Yolo, SSD - same result. The rectangles are not perfect. If you look at examples available online, you will see exactly what I mean: look at the image In a picture this article provides you can see that the proposed rectangle "cuts" the object, leaving part of it outside, while in some other places leaves too much space between the object and the rectangle.

My question is: how can I get an EXACT rectangle, providing my training data are accurate? Any fine-tuning tricks, different approaches, links - anything will be appreciated. Thank you.

Answer 1

One possible way of doing this (goes without saying, but this is just one approach, no idea if this will suffice for your particular problem)

Essentially turn to classical techniques as a post-processing step after your NN step.

• Get the approximate locations of the coins from the network.
• Over-crop the bounding boxes - just increase the width and height of all bounding boxes by a fixed amount so that the full coin is likely to be within the BB at all times. Obviously you will run into instances where you will do this do this even when you didn't have to crop. But that could be okay depending on your application. The idea is to always have a BB covering the entire coin at all times, even at the risk of over-cropping some images.
• Create new smaller images from the bounding boxes. Hopefully, this will contain images of just single coins.
• Create a mask of the pixels likely containing the coin using a simple color threshold (or alternatively also use something like comparing with a HOG description of the coin). Since the images have already been cropped with the help of the NN, this might be relatively easy to do and won't have a lot of false predictions.
• Do hough circle detection on the masked out image.

This method can likely fail if you have overlapping coins and coins cluttered way to close in the same image. But with the help of Hough circle detection, it should be robust to some amount of cluttering. You can also do some filtration on the circles after the last step.