QGeoAI logoQGeoAI

Model Configuration

The Model section allows you to define the task, architecture, and model parameters based on the imported dataset. The plugin is evolutive, and new tasks, architectures, or backbones may be added in the future.

QGeoAI QModel Trainer model configuration

Task

  • The task is automatically pre-selected based on the imported dataset type.
  • QModel Trainer supports the following tasks:
  • - Semantic segmentation: pixel-wise segmentation to identify areas of interest in rasters.
  • - Instance segmentation: segmentation of individual objects with separate masks for each instance.
  • - Instance segmentation (YOLO): faster than Mask R-CNN but less precise.
  • - Detection YOLO: object detection and localization using bounding boxes.

Architecture

  • Available architectures depend on the selected task:
  • - Semantic segmentation: unet, unet++, manet, linknet, fpn, pspnet, pan, deeplabv3, deeplabv3+
  • - Instance segmentation: mask rcnn
  • - Instance segmentation (YOLO): YOLOv11-seg
  • - Detection YOLO: YOLOv11, YOLOv11-obb

Backbone and Pretrained Weights

  • For compatible tasks and architectures, select the desired backbone.
  • You can choose whether to use pretrained weights to initialize the model.