How to interpolate ground-truth from sparse measurements? (about figure6 in the paper)

[keishatsai]

Hi,

I am wondering that how to interpolate ground-truth from sparse measurements as it says in the paper.
There is no explanation for that.
Could anyone teach me how? Thank you :)

[joschuck]

This is a quick solution. Just apply a max convolution. In this example I read a grayscale .png file, convert it to a numpy array, then apply the max convolution.
See this gist for smoothing lidar images
I hope I understood your question.

[keishatsai]

@joschuck
Thanks for the help. This is exactly what I am looking for. Let's me try the method first.

[keishatsai]

Hi @joschuck,
I just found out there is another problem. Have you tried to reproject the 3D velodyne points to the colour image because I just realised that there is no groundtruth in some KITTI dataset, i.e. 2011_09_26/2011_09_26_0009_sync/

What would you suggest me to do? Just reproject the velodyne points?
Thanks.

[joschuck]

Yes I've had this problem too. For some images there is the ground truth missing and the other way around. I just ignore those. I.e. change the following line 96 (commit 7b98ef2) in kitti_eval/depth_evaluation_utils.py to

if os.path.isfile(data_root + im) and os.path.isfile(data_root + vel):

[keishatsai]

Hi @joschuck, I was trying your code, but errors come in.
It seems that I have to form the image first from the velodyne bin file, and I right?

[keishatsai]

Hi @joschuck,
I got error message when running your demo code. Could you help me to clarify the problem?
Thanks
I named module Image from PIL as getVELOIMG

Traceback (most recent call last):
File "projectVelo_detail.py", line 238, in
smooth_lidar_image = getVELOIMG.fromarray(ndimage.filters.maximum_filter(lidar_image, (5,5)), mode='I')
File "/usr/lib/python2.7/dist-packages/scipy/ndimage/filters.py", line 970, in maximum_filter
cval, origin, 0)
File "/usr/lib/python2.7/dist-packages/scipy/ndimage/filters.py", line 903, in _min_or_max_filter
sizes = _ni_support._normalize_sequence(size, input.ndim)
File "/usr/lib/python2.7/dist-packages/scipy/ndimage/_ni_support.py", line 64, in _normalize_sequence
raise RuntimeError(err)
RuntimeError: sequence argument must have length equal to input rank

[tinghuiz]

For interpolation I just use the depth value from the nearest pixel with a valid measurement.