AI4PublicPolicy / AI4PublicPolicy Policy Extraction Demonstration

AI4PublicPolicy Policy Extraction Demonstration

In this video, Petro Gnatyuk from GFT Italy presents the Policy Extraction demonstration of AI4PublicPolicy. The demo is based on the Lisbon pilot of the project, focusing on one of its districts to estimate the number of installed solar panels using aerial images and a trained AI model.

In this demonstration of the AI4PublicPolicy project, the Policy Extraction component of the VPME is presented, based on the Lisbon pilot of the project, aiming to estimate the number of installed solar panels using aerial images and a trained AI model.

The Policy Extraction Component employes a process flow of 4 different phases: (i) Data preparation, (ii) Training, (iii) Evaluation and (iv) Inference phase.

In the Data preparation phase there is:

  • A Label Converter component responsible for extracting labels from the datasets and converting them to the format that is supported by the yolo5 framework, then saving these labels as a separate txt file for every image.
  • An Image Splitter responsible for splitting big 5k or 6k pixels images to smaller (1024 pixels) as required for the training algorithm.

Having prepared the data, in the Training phase, the actual training of the model takes place in the following steps:

  • Getting all necessary yolo5 source files and then configuring the dataset by specifying all necessary paths and class names for object detection and saving the file.
  • Image download (both training and validation datasets) to start the model training. This process takes several iterations. The output of the previous iteration is used as a starting pre-trained model for the next iteration to achieve optimal accuracy with the same amount of data.

In the model Evaluation there are 2 metrics:

  • An obj_loss which shows the error rate for the object detection. With every iteration the error numbers get lower.
  • The box loss that shows how close the box is to the detected panel.

Finally, in the Inference phase, a data management API is used to retrieve images for one specific Lisbon region and run the inference process to detect and to estimate the number of panels in that region.


For more information please have a look at the demo below