Train a model

Menu: AI /  Train a model

The Train a Model feature introduces a complete, end-to-end workflow for creating and improving AI classification models directly inside VisionLidar. This new workflow is built around two key concepts:

1. DNN Projects – A structured container for training and validation data

2. VisionLidar 2.0 Engine – A next-generation AI engine designed for speed, efficiency, and accuracy

This guide walks you through the full process, from project creation to training results analysis.

Before training a model, you must create or open a DNN Project. A DNN project defines:

• The training and validation datasets

• The class definitions

• The configuration used to train one or more models

1. Click Create Project.

2. In the dialog:

   • Select the folder location where the DNN project will be saved.

   • Select the PLY files to be used for training.

⚠️ Important

• Only PLY files generated by VisionLidar are accepted.

• PLYs created by other software are not supported.

Once created, the project becomes the active context for model training.

Open

After project creation, the selected PLY files appear in the Preview tab.

• Files can be assigned to:

  • Training set

  • Validation set

• Use drag and drop to reorder files or move them between Training and Validation.

This step is critical, as the validation set directly drives model evaluation and early stopping.

💡 A high-quality validation set is essential. It must:

• Be representative of the full dataset

• Contain all relevant classes

• Avoid being too simple or incomplete

Poor validation data leads to misleading accuracy and weak models.

For each PLY file, you can:

• Edit class names

• Modify class IDs

• Change class colors

💡 These edits must be done per file, ensuring class consistency across the dataset.

PLY files can be displayed using:

• Class colors

• Intensity

• RGB (if available)

This helps visually verify labels and data quality before training.

If you forgot to include files or want to expand the dataset later:

• Use Import Files to add more PLYs to the same DNN project.

• This allows reuse of a project to train improved or extended models, for example by adding more fences or scenes.
  
  

  Open

  

Once data verification is complete, switch to the Training tab.

Open

You must choose the engine used for training:

• More than 10× faster

• More than 10× more GPU-memory efficient

• Higher model accuracy

• Can be trained on an average PC or laptop in a few hours

• Supports multi-GPU training

• Legacy engine

• Same parameters as VisionLidar 2.0

• Does not support multi-GPU training

➡️ All parameters described below apply to both engines unless stated otherwise.

Defines how many times the model iterates over the dataset.

• More epochs generally improve learning

• More epochs also increase training time

Early Stopping

• Training stops automatically if no validation improvement is observed for 30 consecutive epochs

• The best model is saved automatically

Automatically adapts batch size and worker count based on available GPU memory.

Use this mode to create a brand-new model.

You can train using:

• Geometry

• Intensity

• RGB

• Any combination of the above

To simplify setup, predefined profiles are provided:

• Mobile

• Indoor

• General Aerial

• High-density Aerial

If your data does not fit these profiles, select Custom and define parameters manually.

Subsampling Distance

• Minimum feature size recognized by the model

• Should be ≥ average distance between points

• Recommended: 10–30% larger than average point spacing

Neighborhood Radius

• Size of the spherical neighborhood used as model input

• Should be ≤ half the size of the fence (assuming a square fence)

• Recommended: 60–80% smaller than the smallest PLY size

These parameters strongly influence:

• Model generalization

• Training stability

• Final accuracy

Correct values ensure that each epoch sees a slightly different representation of the data, improving robustness.

Use this mode to improve an existing model.

You can:

• Continue training with the same PLYs

• Train using new or expanded datasets

A new capability allows:

• Adding classes to an existing model

• Removing classes from an existing model

⚠️ Important rules:

• PLY files must contain the classes you want to add

• Classes you want to remove must not exist in the PLYs

Class changes are managed through the Classes List section.

When you click Start, training begins immediately and the interface switches to the Results tab.

The graph displays:

• Training mIoU (blue)

• Validation mIoU (orange)

• A progress bar tracking training completion

• Smooth increase in both curves indicates healthy learning

• Large gap between training and validation curves indicates overfitting

• No validation improvement for 25 epochs triggers automatic early stopping

• Sharp oscillations suggest insufficient or inconsistent data
  
  

  Open

  

The Log tab records:

• Training steps

• Warnings

• Errors and failures

If training fails or behaves unexpectedly:

• Send the log file to the VisionLidar support team

• Logs allow rapid debugging and precise issue resolution

The Train a Model feature transforms VisionLidar into a complete AI training environment:

• Structured DNN projects

• Flexible dataset management

• Powerful VisionLidar 2.0 engine

• Intelligent early stopping and performance tuning

• Clear training diagnostics and results visualization

With careful data preparation and validation selection, users can now train accurate, production-ready AI models in hours instead of days.