Generating accurate floor plans from 3D scans

DigitalBridge builds high-tech applications to help our users embark upon their renovation projects, allowing them to visualise what their room will look like after a re-decoration, and giving them a realistic idea of what they can do with their space. We are using latest developments of Augmented Reality and 3D capture devices to offer customers new possibilities for their next home renovation.

In this blog post we are going to talk about a technique under development to estimate a floor plan automatically from a 3D scan of a space, and outline some of its uses.

Let’s start with the definition of what we mean by a floor plan. A floor plan is the architectural fingerprint of a room, giving its overall dimensions and the positions of key structural features such as doors and windows. It is a necessary first step if you want to start planning a substantial renovation of your kitchen or bathroom.

Creating a floor plan typically requires you to get a measuring tape, and take detailed measurements of the length of each wall, as well as the height, width and position of each window and door. All this before you have even started thinking about a new design! Our user research has shown that this is a major impediment to customers beginning the journey of planning their new kitchen or bathroom.

Our goal is to simplify this process for users by allowing them to use their mobile devices to capture the 3D layout of their room, from which we can automatically compute a floor plan. We hope that this technology will make it easier for our users to begin planning their next renovation.

 Example floor plan,  via Wikimedia Commons

Example floor plan, via Wikimedia Commons

Method

Recent consumer devices based on 3D depth cameras, ARKit or ARCore, allow you to capture information about the real world by moving your device around. We want to use the data captured from these devices to automatically create a floor plan. The method used here requires a dense point cloud or a triangulated mesh of the input space. This can be obtained by either a specific scanning equipment or a device able to capture depth, such as commercial 3D depth cameras. Our tests are conducted with Google Tango, a surprisingly compact device which easily enables the user to quickly scan a space.

We are currently testing the method with the more recent ARKit/ARCore technology, in order to offer this feature to an even broader range of customers.

 Google Tango Development Kit

Google Tango Development Kit

The current algorithm attempts to estimate the geometry of the room starting from the walls. We classify points that are likely to belong to walls and extract 3D primitive candidates. We extract in a similar way floor and ceiling and a final optimisation procedure creates the most likely valid room with a closed topology.

Given the structure and dimension of walls, floor and ceiling, the output of our algorithm can be either a 2D floor plan highlighting walls or a 3D simplification of the room.

 From left to right: Input Scan, 3D CAD Model, 2D Floor plan

From left to right: Input Scan, 3D CAD Model, 2D Floor plan

The method is robust to various room types and shapes, and works with both Manhattan (walls orthogonal to each other) and non-Manhattan rooms.

Some of the benefits of this method are:

  • Accuracy: the error is less than 5 cm, and even smaller if the scan is high quality. The more accurate is the scan, the more precise is the output.
  • Speed: given a 3D scan of a room, our algorithm computes the floor plan in a handful of seconds. Current handmade techniques require human input and significantly more time and effort.
  • It will be available as a web service with an API, so that customers and retailers will be able to integrate it quickly in their own pipelines, similarly to what we have done with our main product.

Conclusions

ARKit and ARCore are the next steps in extending this technology to more devices, and we are currently under active development to being able to support them.

At the current stage, we are working on adding to the representation other architectural features that are less likely to be moved in a redecoration work, such as doors, windows, pipes or columns. Bringing such elements in the virtual space would give more expressive power to the synthetic representation and help the user design with a more faithful model of their space, thus improving the overall experience. We are also considering extending the approach to deal with an entire apartment automatically, to extend even further the possibilities of the system.

We envisage the whole system to be a fully automatic scan to CAD application, where a user scans the space and the algorithm converts it to a high quality emptied virtual representation, useful for more advanced redesigning of space and AR/VR design experiences. This technology, combined with other DigitalBridge products, will help the user with the rather daunting task of redecorating a space.

Antonia Lock