How to prepare the datastore for point cloud classifications

Hi everyone.

I want to use Deep Learning for point clouds' classification starting with at least one of extensions I have available (.ply, .bin, .csv, or .mat). I found this code https://it.mathworks.com/help/vision/ug/train-classification-network-to-classify-object-in-3-d-point-cloud.html that might be useful to me but i have some problem with the creation of the datastore. In fact, I can't insert the labels correctly and the code already fails at the "preview (dsTrain)": the second box is not "1x1 categorical" but undefined. My point clouds files are 32x32x32 cell, how can I prepare correctly the datastore?

Thanks

Answer 1

[Shashank Gupta]

Hi EMMANUELE,

Currently lidar datastore can be formed in the same way other datastores are form through fileDatastore function, I am not sure what extension you are trying to import but .csv, .mat or .ply can easily imported through csvread/readTable, load or pcread. For .bin file you may explicitly need to write a custom reading function in fileDatastore to import. Even Labels are handled in the same way all datastore in MATLAB does, it iterate over all the subfolder in the given folder and assign a category to each one of them. Refer to the same fileDatastore function to understand how label is imported in general. You can also refer to this which is one of the example implementation for lidar point cloud importing with labels.

Hope it helps.

Cheers

[EMMANUELE BARBERI]

Hi, thanks for your help.

I modified something in the function you showed me and now I obtain the predefinite form of datastore {(pointCloud) (label)}, but at line 5 ( the link I sent: "labels = readall(dsLabels);") Matlab gives me the error in the picture and I'm unable to solve it. Some advice?

Thanks