An augmented-reality previewer that puts real tile designs on a customer's own floor

Client Industry

Retail · home improvement & interior finishes

Key Challenge

Buyers cannot picture a full floor from a sample tile

Core Tech Stack

Computer Vision · AR rendering · Segmentation · Mobile

Deployment Model

On-device AR over live camera or photo

SoftmaxAI has developed an augmented-reality tile previewer which projects the designs from the catalogue onto a customer's own floor. It uses a computer-vision process to segment the floor surface in either a live camera feed or a static photograph and to estimate the plane together with its perspective; the rendering stage then warps the chosen tile texture onto that plane at the correct physical scale, properly placing the grout lines and transferring the lighting from the scene onto the rendered surface so that the final result appears to be part of the room. The different designs can be switched over instantly, compared side by side, and shared — thus turning the process of selecting tiles into one based on evidence rather than on imagination.

• Small samples mislead at scale. The pattern, tone, and repetition appear and behave very differently when viewed over an entire floor compared to when a single tile is held in the hand — this is the reason why buyers end up regretting the decisions they make on the basis of samples.
• Showroom light is not room light. A tile assessed under retail lighting looks different in the room it is intended for, and no description is capable of resolving this.
• Trial and error is slow and expensive. It takes days to visit stores, take the samples home, and return them, yet uncertainty remains at the end.
• A convincing overlay is a hard vision problem. Rendering a tile onto a photo is only useful when the floor boundary, perspective, scale, and lighting are correct — if any of these aspects are inaccurate then the preview appears as if it has been pasted on and is therefore ignored.

The previewer functions as a pipeline that runs on the device itself; it segments a camera frame or a stored photo in order to isolate the floor, estimates the plane and perspective of that floor, warps the selected catalogue texture onto the plane at true scale with the scene lighting being applied, and then returns the composite view for the purpose of comparison and sharing.

• Spec before code. We decided upon the things that the preview had to get right — namely, the floor boundaries, the perspective, the tile scale, the grout, and the lighting carry-through — and set those criteria as the conditions for acceptance before implementation began.
• Segment the surface first. A vision model isolates the floor plane in the scene, so the rendering layer knows exactly which pixels it owns and where the walls, furniture, and occlusions begin.
• Estimate the plane, then warp to it. The perspective is obtained from the detected plane and the texture from the catalogue is warped onto it at true physical scale, so that the tile size in the preview is the same as the tile size in the box.
• Human-owned visual quality. Reviewers compared rendered previews against photographs of real installed floors and tuned lighting, scale, and grout treatment until the output was trustworthy enough to buy from.

• The decision moves to the room. Customers evaluate tiles in the space they are actually tiling, under its own light, instead of extrapolating from a sample under showroom light.
• Comparison becomes instant. Since only a single tap is required to switch between designs, the shortlist is actually compared rather than remembered one sample at a time.
• Live camera or an existing photo. Whether you point the phone at the room or load in a picture that was taken earlier, the same pipeline can be used so that a decision can be revisited without having to go back to the space.
• Shareable choices. Rendered previews are saved and shared, which is how a decision involving a partner, a designer, or a contractor actually gets made.
• Extends beyond tiles. The same scene-understanding and rendering approach carries to wall paint, furniture, and other interior finishes without rebuilding the pipeline.