Minimize false positive results doing some image processing in an object detection model

Question

I am using Mask-RCNN to solve an object detection problem. This is an implementation of Mask R-CNN on Python 3, cv2, Keras, and TensorFlow. I am trying to identify the damaged area of a truck. The results which I got are good when I am running the model on those images which do not have any shadow or reflection from the surrounding. But the model fails on those type of images which has a shadow or some other reflection. I have used some image processing techniques which are 1. Converting images to grayscale and 2. Color processing. But both of them not given any good results.

Please suggest what i can do to minimize false-positive results.

Answer 1

The problem with training custom classifiers is that even if you have enough images of the object itself, there isn't enough data of that that same object in different contexts and backgrounds.

I'd suggest you to augmentate the data by applying some sorts of distortion, including artificial shadows and reflections. By doing this, you will get more data with different contexts and minimize your false-positive results.

There are several tools for doing this. One of them is albumentations : https://github.com/albumentations-team/albumentations. It allows you to create a lot of image augmentations, including random shadows.