BauFoto Image Duplicate Finder

Figure 1: Layout of the BauFoto Tool
Figure 2: Two expert systems for duplicate detection work together.
Figure 3: Details on how both expert systems detect duplicates.

Inspiration

Finding Images with duplicate content is a mundane but important documentation task for civil engineers. Digitalization and AI can help with that!

What it does

BauFoto is an Image Browser for duplicate finding. A user uploads photos, coordinates where the photos were taken, and a list of objects identified in the photos by an AI algorithm. A user adjusts the parameters of the two duplication-finding algorithms working together to achieve optimal results. The user saves the identified duplicate images to the desired location for further processing. Figure 1. shows the layout of the tool.

How we built it

BauFoto consists of two expert systems working together (see Figure 2.): A rule-based expert system identifies images with overlapping fields of view (triangle-shaped) based on the information about the coordinates and orientation of the photos (it is given from a separate source as a list of coordinates and orientations). An AI-based expert system uses the output of the object detection model to identify images with similar characteristics (e.g., images that both have an object of a particular class or lack thereof.). Each of the two expert systems returns a list of potential duplicates. The final result is the intersection of two results. The rule-based system has user-controllable parameters for the size and shape of the fields of view. The AI-based system has user-controllable parameters for the objects to consider in the image and the type of comparison required to define a duplicate (e.g., the existence of an object match, the exact number of objects match). (See Figure 3.)

PyQt for GUI, Pytorch for generation of list of identified objects, matplotlib for plotting, Python for everything else.

Challenges we ran into

The format of the coordinates was not compatible with the standard libraries. We needed to preprocess it. We started with a Flask as GUI but run into problems, on Saturday evening we pivoted to PyQt as it was a technology we knew better.