I am training pointCNN in ArcGIS pro 2.7.1 and Python 3.7.9. I am using synthetic point clouds, which I made using photogrammetry. I can run pointCNN, but it is not classifying points in my test dataset. If I run point.cnn.compute_precision_recall(), it shows that the algorithm converged on one class.
if I change the hyperparameters, I cannot improve the result, I can only switch the result:
Does this indicate that pointCNN is not running properly? or that I need more training data?
For reference, here is my script:
Also, I'm curious if this problem is related to how I created my training dataset. I created training and validating datasets using the following steps. 1) I segmented the point cloud in CloudCompare and I exported these segments as .LAS. 2) I imported the .LAS point clouds into ArcGIS pro by creating a LAS dataset and adding the file. 3) I classified all the points in each segment as a single class- I used class 2 and 6 even though I am not using a point cloud that includes buildings and ground. 4) I exported each classified segment using the "Extract Las" function, this gives me a .las and .lasx files. 5) I created a train and val directors, each directory contains 4 files: 1 .las and 1 .lasx for the class 2 and 1.las and 1 .lasx for class 6.