How automatic object detection in large point clouds can be efficient and scalable
Automatic object detection in large point clouds becomes efficient and scalable when not only automated and customer-specific classifications of point clouds but also tailor-made neural networks are implemented. For this we offer a new new AI Training Service, which significantly reduces he effort of manual annotation.
What’s needed to start with this Service is training data for the neural network, which is the common bottleneck for any AI approach. When there is a decent amount of training data our data scientists can start training the model. To measure ‘decent’ can vary a lot depending on your use case and the desired performance, but we recommend 5–15 days of labelling work to provide a realistic start for a Proof of Concept. That means you will get a solid estimate of the future potentials of AI but not yet a production-ready model.
To generate this data, the point clouds need to be uploaded into Pointly and manually classified. You can either book a Labelling Service provider from Pointly’s Partner Network at competitive rates to take over this task or classify the data yourself in Pointly. Pointly was developed to make this work as fast and easy as possible through intelligent selection tools, a user-friendly interface, and smart segmentation. If you already have classified data, you can leverage this as training material for your classifier and get a jump start with the Service.
After the initial training, the model can take over classifying the point clouds automatically. The user can then manually improve the model outputs inside Pointly to “teach” the model to get even better. The more often this process is repeated, the more robust the model gets. So do the manual work a few times and then rarely ever again.
Our team works with an iterative approach during the training process. That means that the data is split into 3 evenly sized sets, beginning training with the first, then adding the second and so forth. This allows Pointly to evaluate how performance increases with additional training data and it gives you more flexibility and control when generating training data. To measure performance we ask you to choose a test site to be put aside first of all. This test site will not be used for training but only for evaluating the final performance of the neural network. During the training process on the separate training data, we will provide you with regular updates on the performance, based on cross-validations within the current training set.
Overall, the AI Training Service includes:
— A Professional Pointly subscription to upload, manage and classify your point clouds.
— A personal consultation session with us to assess your goals and define KPIs.
— Regular performance updates on the training progress of the model.
— An AI point cloud classifier (that is the neural network classifying your data).
After the Service ends you can keep using your AI point classifier for classification tasks as well as re-train it with more data based on standard charges.
You can inquire the AI Training Service on Pointly’s website via a form here.
Pointly — Point out what matters
