Evaluating Vision-language Models for Zero-shot Room Area Estimation in Floor Plan Images

Abstract

Vision-language models (VLMs) have shown strong performance on multimodal reasoning tasks, yet their ability to perform quantitative analysis of technical drawings remains largely unexplored. This study evaluates the instruction-guided zero-shot performance of three open-source VLMs, on the task of estimating room areas from color-coded residential floor plan images. Of the three models tested, two produced quantitatively evaluable structure outputs. A dataset of 100 floor plans, comprising 472 plan-level room-type evaluation comparisons derived from 894 ground-truth room instances, was used to compare predicted and ground truth areas. Results show moderate predictive correlation (R² ≈ 0.71), but substantial estimation errors (MAPE ≈ 43%), with noticeably larger errors for smaller rooms. A classical pixel counting baseline, using color segmentation, achieved MAPE = 5.68% at the room-instance level, and MAPE = 1.58% when evaluated at the floor plan level, highlighting the limitations of the tested VLMs for precise geometric estimation under this controlled setup. These findings suggest that while the evaluated open-source VLMs can approximate spatial proportions in simplified color-coded layouts, their quantitative accuracy remains significantly below that of deterministic image analysis methods.