Understanding Ground Truth Label

[newTypeGeek]

I am able to train the Structured Forest using the BSDS500 image dataset. But I am confused with the ground truth label in this directory BSR/BSDS500/data/groundTruth/train.

For each image, the groundTruth label in .mat format consists of 6 struct with 2 fields: (Segmentation and Boundaries).

Question 1:
I don't understand why we have 6 labels for each image. Isn't it more natural to have one label (one struct) per image?

Question 2:
And also, If I want to train my own dataset, how can I setup my own ground truth label (both Segmentation and Boundaries)?

Thanks.

[CNHNLP]

@newTypeGeek I have the same question.Have you understand it?

[dengdazhen]

@newTypeGeek I have the same question.Have you understand it?

Each struct is an annotation by one annotator, including the segmentation and boundaries. You can refer to the paper A Database of Human Segmented Natural Images and its Application to Evaluating Segmentation Algorithms and Measuring Ecological Statistics for more details.