Hi Daniel,
Great to see the custom training being used for railway applications! :)
How do your training and validation files look like? You would need to use the complete point cloud for training and validation. Using an isolated region (even including a 20m radius) will not be enough as your model needs to see the object of your interest in real-world context. 250 masts sounds impressive!
Please try retraining using complete point clouds (trained region + all other points).
What is your voxel size ?
Feel free to contact me at khrystyna_bezborodova@trimble.com
------------------------------
Khrystyna Bezborodova
------------------------------
Original Message:
Sent: 01-18-2024 07:14
From: Daniel Hermann
Subject: own point cloud classification using deep learning (based on Tensorflow)
Hey guys :)
I have the following problem: I want to use the option implemented in TBC to create my own classes in a point cloud to perform object recognition with deep learnig. I work for a railroad company and 5 different types of lattice masts are to be recognized on the railway.
In the first attempt, I cut out all the masts separately and divided them into test and validation files. The accuracy was bombastic, but only for the isolated poles. When I applied the model to a normal point cloud, the result was unusable.
In the second attempt, I added the corresponding point clouds in a 20m radius to the masts. Thus, the files contain 20-meter islands in a 300km area. Again, the result is not usable.
Has anyone had any experience with deep learning training and can give me tips and suggestions on how to achieve the desired result? My ratio of test and validation size is 3:1 and I use a total of almost 250 masts. The files are in .las format.
Best regards,
Daniel
------------------------------
Daniel Hermann
------------------------------