Hello Kamesh Sonti ,
your question is somewhat vague. But I'll try to help with my answer.
First of all, one of the things you may want to do is measure how similar (or dissimilar) your segmented image vs ground truth are. To this end, you may check this chapter in which several metrics are well explained:
https://pdfs.semanticscholar.org/1453/171acb23daae8e4075dc741dc21c6ad89800.pdf
However, I've seen some other metrics related to the image segmentation tasks, so perhaps you want to search for the following:
- Rand Index: Measures the similarity between two clusters.
- Variation of information: It is similar to mutual information; it is a metric that gives a measure of distance between two clusters.
- Global consistency error.
Hope it is helpful for you.
In general, image segmentation is a difficult problem and this is due to the gap between our sematic understanding of the concepts that were captured by the image and their low level representation in terms of color, texture, motion, etc. Mathematically, the image representation of the real world is an ill-posed problem and cannot be easily constrained. There are many statitsical ways to judge the performance of image segmentation but this does not imply there correctness from a semantic point of view.
Having said that, this is does not deny the fact succesful segmentation performance can be acheived in controled applications and environments such industrial objects segmentation.
A more concervative approch is to generate manual results of segmentation to serve as benchmarks for the ones that will be generated by the computer-based algorithms.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images