I have a lot of duplicate images. I need to make a selection to reduce the amount of images the Mask RCNN model will perform inference on.

In every collection of duplicates, the images slightly differ. One is brighter, darker, different shadows, different reflection, etc. But all duplicate images are essentially capturing the same object(s) from the same angle with the same background, etc.

What computer vision methods can help me to know before inference what image would produce the most results after inference?

I have tried:

  • edge detection (cv2.Canny followed by sum()): Selecting the image with the most edges. In my use case there does not seem to be a strong predictive relation between amount of edges and amount of objects detected.
  • ...
  • $\begingroup$ clarify what you mean by "produce the most results after inference?" $\endgroup$ Mar 2 at 20:09
  • 1
    $\begingroup$ to get an idea of the scale, how many images are there? is it a public or private dataset? did you try any of techniques mentioned here medium.com/mlearning-ai/… $\endgroup$ Mar 2 at 20:10
  • 1
    $\begingroup$ What I mean by "produce the most results" is that the model would detect the highest amount of objects. The model does not always detect the same amount of objects for every one of the duplicates. It seems that certain properties of the image increases the amount of detections that the model makes. The dataset would in total contain more than 100 000 images and it is private. I haven't tried them. Thank you for the link. Will look into that! $\endgroup$ Mar 6 at 14:16


You must log in to answer this question.