When Should a Network Emit Geometry, and When Should It Detect It? Readout, Reconciliation, and Representation in Floorplan Vectorization
This paper compares autoregressive sequence emission versus heatmap detection readouts for floorplan vectorization, finding that detection generally outperforms emission on real-world scans while sequence decoding excels on clean vector renders, ultimately demonstrating that deterministic fusion of both approaches yields the highest accuracy and that output representation matters less than previously assumed when reconciliation and matching training recipes are applied.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where every building plan, from a cozy apartment to a sprawling office, exists only as a flat, printed image. To architects, engineers, and navigation systems, these static pictures are useless. They need a living, editable map made of lines and shapes that a computer can understand and a human can modify. Turning a flat drawing into this digital structure is a task known as vectorization. For years, the industry has debated the best way for a computer to perform this translation. One school of thought believes the computer should act like a scribe, writing out the walls one by one in a long, ordered list. Another school argues the computer should act like a surveyor, first spotting every corner and line on the page, and then connecting the dots to build the structure. This question of method—whether to write the geometry or to find it—has been the subject of intense theoretical discussion, but until now, it lacked a clear answer based on real-world performance.
A new study by independent researcher He Zhang cuts through the theory by testing both methods on the exact same computer brain. The researcher trained a single artificial intelligence network to understand floor plans and then asked it to produce the final map in two different ways. In one scenario, the network wrote the walls as a sequence of coordinates, much like a person tracing a path with a pen. In the other, the network looked at a heat map—a visual guide showing where corners and lines were most likely to be—and assembled the walls by connecting those points. The results revealed that there is no single "best" way to do this. Instead, the right choice depends entirely on the size of the building and the quality of the drawing.
When the computer was asked to read real-world scans of existing buildings, which often contain noise, smudges, and complex details, the surveyor approach proved superior. By detecting corners and lines first, the system produced more accurate walls, especially for large, complex floor plans. The improvement was significant: on large plans, this method reduced errors by a measurable margin compared to the writing method. However, the story flipped when the computer looked at clean, computer-generated drawings. In these pristine environments, the writing method, which had been trained on similar styles, outperformed the surveyor approach. The study found that the size of the plan mattered greatly; for small apartments, the writing method was often better, but as the plans grew larger and more intricate, the surveyor method took the lead. This suggests that the difficulty of writing a long, precise list of numbers increases with the size of the task, while spotting patterns on a dense image remains stable.
The research also challenged a long-held belief about how these maps should be built. Many experts assumed that if a computer tried to draw entire rooms first and then derived the walls from them, the result would be topologically flawed, leading to double walls or gaps. The study tested this by comparing a system that built rooms first against one that built walls first. After applying a standard correction step to fix minor errors, the room-based system performed just as well as the wall-based system. The only major difference was that the room-based system sometimes missed walls that did not enclose a specific room, such as a freestanding partition or an exterior boundary. This was not a failure of the method itself, but a limitation of what the room-shape format could express. The study concluded that the choice of output format matters far less than previously thought, provided the system includes a step to reconcile and clean up the final lines.
Perhaps the most practical finding was how these two different approaches could be combined. The researchers discovered that feeding the output of one system into the other as a hint did not help; the computer gained no new information from a second guess at the same image. However, merging the final results of both systems after they had finished their work created a much stronger outcome. By taking the best walls from the room-based system and adding the missing walls detected by the surveyor system, the researchers created a hybrid model that was significantly more accurate than either method alone. This combined approach reduced the amount of human work needed to correct the final map by about thirty percent, a crucial metric for professionals who rely on these digital drafts.
The study also introduced a new way to measure success that goes beyond simple accuracy scores. Instead of just counting how many lines were correct, the researchers calculated the "edit cost," which estimates the actual human effort required to fix the computer's mistakes. They found that the hybrid system required the least amount of manual correction, making it the most efficient path forward. To ensure these findings could be tested and verified by others, the researcher released a new, standardized benchmark containing nearly seventeen thousand floor plans, along with the corrected data and the code used for the experiments. This work provides a clear, data-driven rule for the future: for large, messy real-world scans, let the computer detect the geometry; for small plans or clean digital drawings, let it write the geometry; and for the best results, let both methods work together to create a map that is ready for human use.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.