grid_map
grid_map copied to clipboard
[grid_map_cv] Conversion to OpenCV force NaN to 0
When converting a GridMap to an OpenCV image with the method grid_map:GridMapCvConverter::toImage
, the cv::Mat is initialized with zeros (black pixels) (here) then, only finite values are copied from the GridMap to the cv::Mat (here).
In other words the NaN
values are automatically translated into black pixels.
For my use case (processing a map during a SLAM), I have a lot of NaN
values and I need that information in OpenCV.
So maybe, setting the value of the pixels to the mid value between lowerValue
and upperValue
is more interesting than forcing to 0 by default.
Are you using an image with a channel for the alpha layer? There's a round trip unit test that passes for the CV_8UC4
CV image type and I just added a nan-value/transparent pixel here.
If this does not solve your problem, could you please create a PR with a failing unit tests that shows the issue?
Ah no, I forgot to mention that I am using a CV_8UC1
grayscale image, that is why, there is no other solution to translate a NaN
to a mid value between Free and Occupied.
So is there anything you would like to see differently in the behavior of Grid Map or should I close this issue?
What do you think about setting NaN
values to the mid value between lowerValue
and upperValue
?
*if there is no alpha channel, of couse.
I would keep the current behavior, since I feel it's the one with the least assumptions. Your use case is very special, and if I'm not mistaken, you could achieve your desired behavior by setting your empty/nan cells to the mid-value before converting it to the image format.
Yes, of course, it is what I have done.
But for the map_server
for example, a unknown cell is a pixel with value 128 that is why I was surprised that grid_map
did not picked a middle value.
Anyway, as you said, my use case may be special for the grid_map
because I did not managed to find an easy way to convert pixels position to grid_map
coordinates then OccupancyGrid
coordinates. So I have ended up writing my own conversion methods.
Maybe you could add in the doc that the NaN
values are converted to the value 0 for images with no alpha layer.
I have also encountered this problem.
The issue is that the output of SLAM algorithms for unknown area is -1 in occupancy grids, that gets transformed into NaN in Grid maps and finally that gets transformed into 0 in cv::Mat. That causes cv::Mats to make no difference between occupied cells and unknown cells.
I don't think changing NaN cells to a middle value before converting to grid map is a good solution because an occupancy grid is supposed to give a probability of occupancy of each cell and is not supposed to give a value to unknown space.