I need to calculate the accuracy, precision, recall, specificity, and F1 score for my Mask-RCNN model. Hence I hope to calculate the confusion matrix to the whole dataset first to get the TP, FP, TN, FN values. But I noticed almost all the available solutions out there for the calculation of confusion matrix, only outputs the TP, FP, and FN values. So how can I calculate metrics like accuracy and specificity which includes TN as a parameter?.Should I consider the TN as 0 during the calculations ?
Qamar Ul Islam Don't copy and paste Stack Overflow answers here if you don't know the answer. I have already went through that answer before. At least read the copied content before posting it.
Dear Hasindu Yahamapth Dias Dahanayake ,
Look the link, maybe useful.
https://github.com/matterport/Mask_RCNN/issues/2007
Regards,
Shafagat
Shafagat Mahmudova I have already gone through it ,but unfortunately TN is not calculated there also.
If you have numbers for TP, FP and FN, and everything is measured in pixels, then surely the TN is just:
Size of image - (TP+FP+FN)
But calculating metrics based on the image size might not be the best idea. You could get different results simply by cropping images slightly differently (i.e. your accuracy would change if you just shaved a 1 pixel border off from your sample images - no change in the model, no large visual difference in the sample or ground truth mask). Or you could deliberately select samples with a small target (relative to image size) so that the number of TN is sure to be high relative to all the others - high "accuracy", but no guarantee that you actually managed to generate the correct mask.
Something like Intersection over Union could be calculated using only TP, FP and FN. That might give you a more rigorous metric for your application. You should check the literature to see if this is more common. This paper Article Visual Object Tracking Performance Measures Revisited
might give you a start (it's old, but segmentation started to become an important aspect of Visual Object Tracker evaluation around then).
Hasindu Yahamapth Dias Dahanayake Don't ask question here, do your homework and mind you words, if you don't say anything good about others, than don't say anything.
Hope you get it what I want to say.
I am afraid that for general object detection tasks, thee is no such thing as the number of True Negatives. The definition of True Negatives for object detectors would be something like: This is the number of boxes that contain no object and that are correctly not detected. There are infinitely many such boxes and non-detections, so calculating TN here makes no sense. This includes all metrics that depend on true negatives, such as Accuracy and Specificity, and I have not come across any report that mention those numbers for object detection tasks. All the other mentioned metrics do only rely on TP, FP and FN, which are easy to compute and make sense for this task.
Hasindu Yahamapth Dias Dahanayake if you know the values of TP, FP, FN then use TN=Total size - (TP+FP+FN)
- Badges
- Science topic
- Similar topics
- Counseling Psychology
- Confusion