Innovative Information Technology

Innovitech kft.

Node detector development

The key step of plant-reconstruction is to detect the nodes in the image of the grapevine. In the past we used two different types of models for this purpose: the Detectron2 object-detector, which predicts bounding boxes and optionally the centre of the nodes; and the UNet semantic-segmentation model, which predicts the mask identifying pixels belonging to nodes and the background.

From the test of 2024 February we have learned that different conditions of light and shapes in the background interfare with our prediction. Based on this experience we improved our Detectron2 model in the following ways:

  • generating more precise ground truth data for training,
  • augmenting the training pictures with the aim of making the model more robust to different light conditions,
  • fine-tuning the parameters and hyperparameters of the model,
  • tiling the images for increasing resolution.

With the new model we were able to generate more accurate results both on the training package and the new test images made in February. The value of all indicators of performance (precision, recall and correctness of bounding boxes) were increased.