Decimating pointcloud to 8192 points

[sitagy]

Hello,
I have a pointcloud which about 40,000 points. Since I don't have labels in my data set. I am directly feeding my data to the model in predict.py.
pred_labels = predict_one_input(sess, ops, myData)
I am decimating the pointcloud randomly using tools like cloud compare. Because of random decimation, sometimes the segmentation works and some times it doesn't. Is there a better way to correctly pick points so that segmentation always works?

Thanks in advance