About boundary and mask images

[Ashwin Nair]

Hi,

I had a couple of questions about the boundary and mask images. Kindly refer the attached image for context.

1. Why does the mask image and the area of interest in boundary image have value zero? Shouldn't they be like the label Waterway where the important area has value 1?

2. Why is that the label (in this case waterway) have a smaller area than the area of interest (the green area in the rgb image)? I know there can be overlapping labels but in this example, the only label is waterway.

[sabari nathan]

yes. I think there is a problem in data annotation. Even I am confused with the data.

[Thong Huynh]

Hi, 

Question 2: I think the rest of area will be background

On Thursday, February 13, 2020 at 1:43:11 AM UTC+9, Ashwin Nair wrote:

[agriculture-vision]

Hi Ashwin,

Thank you for your question.

The boundary map indicates that some of the areas are outside the farmland boundary. Those pixels will not be evaluated during testing. In this case, the label map has also been masked out accordingly, it is therefore smaller than the green area in the RGB image.

The Agriculture-Vision Team

[Ashwin Nair]

So let me see if I understand this correctly,

Referring the attached image, the black area in boundary map refers to the area outside the farmland and hence is not considered. The overlapping area between the green area in the RGB image and the white area in the boundary map is the white area in the label i.e. waterway in this example. Is this correct? To be more precise, valid areas are those areas where both boundary pixels and mask pixels are 1 right?

[agriculture-vision]

Hi Ashwin,

That is correct. To be precise, in both the mask and boundary map loaded using numpy.array(PIL.Image.open(field.png)), pixels with value 255 indicate valid regions. Therefore, only regions where both boundary and mask pixels are 255 will be evaluated.

The Agriculture-Vision Team