DeepGlobe Land Cover Classification Challenge Forum

Go back to competition Back to thread list Post in this thread

> Q: how individual class scores are averaged?

If a class is absent on an image, does your metric assign it 0 score if predicted, and 1 score if not predicted on the image? Or you skip this class for evaluation and average across present classes only?
Do you average scores across all 6 classes for an image then across data set?
Can you please give us exact formula or code for metric?

Let's consider an example. An image contains 2 classes. I predict 3 classes. Each class gets Jaccard score 1, 0.8, 0 respectively. Will a metric for the image be
(1 + 0.8 + 0) / 3 or
(1 + 0.8) / 2 or
(1 + 0.8 + 0 + 1 + 1 + 1) / 6 or
something else?

Thank you.

Posted by: rakhlin @ April 11, 2018, 3:01 p.m.

We aggregate and average scores across all 6 classes **all at once** for **all** images in the data set (we do not compute iou for them in each image).

Posted by: jinghuang @ April 12, 2018, 12:26 a.m.

Thank you

Posted by: rakhlin @ April 12, 2018, 1:17 a.m.
Post in this thread