indid.lab - Korean Residential Real Estate Valuation Platform’s View Rendering AI Model Training & Fine-Tuning

There are three main reasons why it is difficult to display accurate views as seen from specific units of condos and apartments. First, for existing condos and apartments, there are tenants occupying each unit already, so it is very difficult and troublesome to ask for their permission to take photos. Secondly, for upcoming projects, since there is no physical building, the only way to provide any idea of a view is through raw GIS modelings. Finally, actual photos and GIS modelings cannot accurately portray the views as seen from each unit due to changes in the environment such as weather, new development projects, time of the day and etc.

We conducted multiple experiments with multiple types of inputs (types of images, elevation, distance to adjacent buildings, scenery and etc.) and different types of AI models we have previously developed for architecture, construction and real estate. The best results we were able to come up with our existing AI models is as shown above. There are multiple buildings in the view, but they all looked different despite being part of a design-coherent complex. The scale is also off as the AI model cannot estimate how far each building is from it’s point of view. The surroundings and the environment is also awkward as it does not know what should be placed at the grey space that is not part of the buildings.

Based on our existing exterior rendering model and landscaping models, we created a completely new AI model that takes into consideration the height at which the scene is viewed from, the distance to buildings that obstruct the view, exterior characteristics of Korean residential condos and apartments. We crawled training images from copyright free photographers’ blogs that specifically deal with Korean residential buildings and trained the new model. On top of that, image processing algorithms were added to classify and filter what type of input images are being handed over to the AI model. The result of the new AI model can be seen above. Accurate and design-coherent representation of exteriors with hyperrealistic details and no errors in the scale or elevation.