I was wondering if anyone had any ideas on how to deal with segments of images you are just not sure whether to label foreground or background?
for example, the data I am working with when I labeled it, I can say I feel confident in about 85-90% of my labelings of foreground and background. However, in the data set there are some regions where I can not be certain where to draw the borders between objects. When I train on this data set and review the predictions, a lot of it seems correct and seems to correspond well with my ground truths, however there are some regions in the original image I seemed to have mis labeled and upon review and it seems that the network somehow learned to correct some of my labeling mistakes! I was wondering how if there is any knowledge or literature on how to deal with parts of your data set you are not 100% confident in your labeling? Particularly when the ground truth is very hard to extract manually. Are there any papers on related topics?
I was wondering if anyone had any ideas on how to deal with segments of images you are just not sure whether to label foreground or background?
for example, the data I am working with when I labeled it, I can say I feel confident in about 85-90% of my labelings of foreground and background. However, in the data set there are some regions where I can not be certain where to draw the borders between objects. When I train on this data set and review the predictions, a lot of it seems correct and seems to correspond well with my ground truths, however there are some regions in the original image I seemed to have mis labeled and upon review and it seems that the network somehow learned to correct some of my labeling mistakes! I was wondering how if there is any knowledge or literature on how to deal with parts of your data set you are not 100% confident in your labeling? Particularly when the ground truth is very hard to extract manually. Are there any papers on related topics?