PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Detection of Mechanical, Electrical, and Plumbing (MEP) Metrics Based on 2D Floor Plans

Tarandeep Singh Mandhiratta, ANK Zaman, Abdul-Rahman Mawlood-Yunis

arXiv:2608.14317v1cs.CVcs.AIcs.HCcs.LG

TL;DR

Extracting MEP information from variable, often fuzzy 2D floor plans is challenging. This paper applies Mask R-CNN to detect lighting symbols, classify light types, and extract associated text, achieving bbox_mAP 0.7596 and segm_mAP 0.7111.

  • Problem

    Variable notation, fuzzy raster drawings, and the lack of standardized MEP symbols make automated floor-plan extraction challenging.

  • Method

    The study applies Mask R-CNN to 2D floor-plan images to detect lighting symbols, identify light types, and extract associated text for power estimation.

  • Results

    bbox_mAP was 0.7596 and segm_mAP was 0.7111, while bbox_mAP_50 reached 0.9850 and segm_mAP_75 reached 0.9219.

  • Takeaways & Limitations

    The approach can automate manual MEP symbol detection, reducing design time and improving workflow efficiency in architecture, construction, and real estate.

Abstract

from arXiv · show

This research developed a neural network-based model to extract various information from 2D floor plans. We detect lighting symbols, identify the appropriate type of light, and extract the associated texts with lights. The study aims to enable efficient floor designing and determining the number and type of lights needed per floor, i.e., allow efficient design and estimate the power requirement of the floor plan. The model was developed using Mask RCNN as the base. The images were annotated and converted into a Coco data format for training the model. The model achieved bbox\_mAP and segm\_mAP values of 0.7596 and 0.7111, respectively. It also performed well at different IoU thresholds, i.e., with bbox\_mAP 50 and segm\_mAP 75 values of 0.9850 and 0.9219, respectively. The developed model will help various industries, such as architecture and construction, to improve design time and create efficient workflows by automatically detecting Mechanical, Electrical, and Plumbing (MEP) objects from floor plans, and it is the first step towards building tools that will help energy-efficient building design.

I. INTRODUCTION · II. LITERATURE REVIEW

The paper motivates automated MEP information extraction from 2D floor plans, addressing notation variability and rasterized drawings through neural-network analysis of lighting symbols, light types, and associated text. Its literature review situates this work within deep-learning approaches for automated floor-plan interpretation and building-management applications.

  • I. INTRODUCTION: Indoor information technologies support facilities management, indoor localization, and indoor data models, increasing demand for databases of indoor information.
  • I. INTRODUCTION: MEP matrix analysis remains underdeveloped despite its importance for efficient engineering design, and this study uses 2D floor plans as fundamental information sources for existing buildings.
  • I. INTRODUCTION: Floor-plan machine learning is challenging because notation varies across institutions and design offices, while archived architectural drawings are often complex, fuzzy raster images.
  • I. INTRODUCTION: The research develops a neural-network model to detect lighting symbols, identify light types, and extract associated text from 2D floor plans.
  • I. INTRODUCTION: After detecting lighting symbols, the method records bounding-box coordinates, saves detected electrical components, and extracts text from cropped symbols to identify component types.
  • II. LITERATURE REVIEW: The literature review describes growing interest in deep learning for extracting MEP components from architectural floor plans used in MEP design and installation.
  • II. LITERATURE REVIEW: Earlier and recent studies applied neural networks, CNNs, FCNs, 3D reconstruction, statistical methods, and Google Vision API to classify symbols and analyze floor-plan features.
  • II. LITERATURE REVIEW: Automating floor-plan analysis without manual intervention is increasingly important for smart homes and building-management systems, despite difficulties in extracting walls, detecting rooms, and preprocessing.

A. METHODS FOR EXTRACTION OF WALLS · 1) Early Techniques for Separating Walls in Floor Plans · 2) Conventional Machine Learning Methods

Wall-extraction methods evolved from manual digitization and rigid image-processing pipelines toward machine-learning approaches that accommodate varied wall notations. Early and patch-based methods remained constrained by drawing-style assumptions, notation diversity, or training-data requirements.

  • A. METHODS FOR EXTRACTION OF WALLS: Overall, wall-extraction techniques advanced alongside computer-vision developments and increased computing power as objectives and challenges changed.The field’s progression spans early digitization, traditional image processing, and conventional machine learning.
  • 1) Early Techniques for Separating Walls in Floor Plans: Early floor-plan digitization converted printed or hand-drawn drawings into CAD manually, requiring time-consuming and expensive tracing.Tracing was performed on computer screens or specialized tracing tables.
  • 1) Early Techniques for Separating Walls in Floor Plans: Initial automated methods targeted patterned drawings with thick, predominantly horizontal and vertical walls, converting them into vectors through morphology, thinning, and line extraction.The supplied passage describes thinning walls to 1-pixel lines for vector representation.
  • 1) Early Techniques for Separating Walls in Floor Plans: Aoki et al. (1996) distinguished internal and external walls using thin and thick lines, then applied line-segment and closed-region extraction with pattern matching.The approach operated on a preprocessed image to identify drawing elements.
  • 1) Early Techniques for Separating Walls in Floor Plans: Ahmed et al. introduced a multistage method involving information separation, structure analysis, and semantic analysis, but assumed walls were the thickest image lines.This assumption restricted the technique to a specific drawing style.
  • 2) Conventional Machine Learning Methods: Architectural drawings lack universal notation: walls may appear as variable-width single lines, parallel lines, or hatching, limiting notation-specific traditional extraction methods.This diversity was identified as a central weakness of conventional approaches.
  • 2) Conventional Machine Learning Methods: Heras et al. introduced a machine-learning wall-extraction approach modeled on Bag-Of-Visual-Words that could manage various wall notations.The method was presented as a response to the limitations of notation-specific techniques.
  • 2) Conventional Machine Learning Methods: Patch-based methods were limited by single-notation training and the need for ground-truth data from diverse image sets.These requirements constrained their generalization across architectural drawing styles.

3) Methods Utilizing Convolution Neural Networks (CNN) … 2) Geometry-Driven Approaches

The reviewed methods span CNN-based junction detection, pixel-level room detection, and geometry-driven room reconstruction. CNN approaches achieved high quantitative results but could fail to preserve wall connectivity, while geometry-based methods depended on accurate wall and symbol information.

  • 3) Methods Utilizing Convolution Neural Networks (CNN): Liu et al. used a simple CNN to predict pixel-level heatmaps for detecting floor-plan corners and crosspoints, or junctions.The method was trained on the LIFULL HOME dataset, containing 1,000 manually annotated floor-plan images.
  • 3) Methods Utilizing Convolution Neural Networks (CNN): The LIFULL HOME dataset was considered sufficiently large for deep learning because it contained 1,000 annotated floor-plan images.The annotations were manual and supported training of the CNN junction-detection method.
  • 3) Methods Utilizing Convolution Neural Networks (CNN): CNN-based approaches achieved high quantitative results but sometimes missed entire walls and could not guarantee wall connectivity.Such failures could make postprocessing-based recovery impossible.
  • B. TECHNIQUES FOR IDENTIFYING ROOMS: Room-detection methods identify and separate floor-plan areas using walls as guidance, so accuracy depends on thorough and correct wall extraction.Errors in identifying walls can affect the resulting room detection.
  • 1) Pixel-level Approaches: Pixel-based room detection analyzes raster images of walls using approaches such as rule-based gap filling and proximity fields for locating potential room centers.Fixed-rule gap filling was limited for complex cases and wall-extraction errors, while proximity fields identified points farthest from walls.
  • 2) Geometry-Driven Approaches: Geometry-based room detection represents walls as vectors and can partition them into nearly convex room polygons by minimizing concavity.Mace et al. used top-down polygon partitioning to fill gaps between walls.
  • 2) Geometry-Driven Approaches: Polygon-loop methods close gaps caused by doors and windows but generally require highly precise prior wall and symbol detection.This precision requirement was described as feasible mainly for low-complexity datasets or with human assistance.

3) Hybrid Approaches · C. PREPROCESSING TECHNIQUES

Hybrid approaches combine structural pixel information with semantic information to detect rooms, while preprocessing methods separate graphics from text and address challenges caused by inconsistent MEP symbols. Proposed solutions include overlap-aware text detection, learning across symbol sets, synthetic data generation, and model ensembles.

  • 3) Hybrid Approaches: Advanced hybrid methods create pixel images of extracted walls and fill structural gaps using information from earlier analysis steps.These methods support subsequent room detection by reconstructing incomplete wall structures.
  • 3) Hybrid Approaches: After gap filling, rooms are detected as connected components, combining semantic and pixel information to balance robustness and accuracy.
  • C. PREPROCESSING TECHNIQUES: Graphics/text separation is a common preprocessing step that detects and extracts text to simplify floor-plan images.
  • C. PREPROCESSING TECHNIQUES: Fletcher and Kasturi filtered connected components using area, aspect ratio, pixel density, and bounding-box dimensions, but failed on overlapping text characters.The method used statistical histogram analysis of technical-drawing features and produced good results while remaining simple to implement.
  • C. PREPROCESSING TECHNIQUES: Ahmed et al. improved graphics/text separation by inferring missing character locations from surrounding detected characters.The approach assumed text characters formed the highest-probability area measure among image connected components when sufficient text was present.
  • C. PREPROCESSING TECHNIQUES: A primary challenge is the lack of a standardized MEP symbol set because projects may use different symbols for identical components.Consequently, generalized detection of every possible MEP symbol across projects is difficult.
  • C. PREPROCESSING TECHNIQUES: Researchers proposed learning from multiple symbol sets and using data augmentation to generate synthetic data.These approaches aim to address variation in MEP symbols across construction projects.
  • C. PREPROCESSING TECHNIQUES: Ensembles of models have also been proposed to increase detection-system accuracy.

III. METHODOLOGY … 3) Image Annotation

The methodology develops a neural-network system to automate MEP metric collection from building floor plans. It proceeds through data collection, PDF-to-image conversion, and manual annotation of lighting symbols and associated text in COCO format.

  • III. METHODOLOGY: The study targets more efficient and accurate engineering design by automating the collection of mechanical, electrical, and plumbing metrics from building floor plans.The system is intended to identify lighting information in floor plans.
  • III. METHODOLOGY: The methodology develops a neural network model through a series of defined steps to achieve automated MEP metric extraction.The methodology is illustrated in Figure 3.
  • A. DATA COLLECTION AND PREPROCESSING: Data collection and preprocessing comprise sequential steps beginning with training-data acquisition and continuing through image preparation and annotation.Figure 4 presents the data collection and preprocessing steps.
  • 1) Data Collection: Because no suitable existing dataset was available, an industry collaborator provided numerous 2D floor plans in PDF format covering different building configurations and layouts.These floor plans served as the training data.
  • 2) PDF to Image Conversion: 300 DPI was used to convert PDF floor plans into images with PyMuPDF, balancing image size and visual clarity for neural-network training.The PDF files were too large to use directly for training, so conversion to image format was necessary.
  • 3) Image Annotation: Manual image annotation was performed with VGG Image Annotator as the third methodological step.The annotations were subsequently converted to COCO data format.
  • 3) Image Annotation: Annotations recorded the locations of lighting symbols and their corresponding text in each image before conversion to the COCO format used for object detection.Figure 5 shows a sample annotated image from preprocessing.

4) Data Splitting · 5) Image Slicing · B. ARCHITECTURE AND NEURAL NETWORK MODEL SELECTION

The methodology split data for evaluation on unseen images, sliced images into overlapping 1000X1000 grids for efficiency, and selected Mask R-CNN for accurate detection and segmentation. The selection balanced Mask R-CNN’s capabilities against its computational cost and slower processing.

  • 4) Data Splitting: The data were divided into training, validation, and testing datasets to evaluate performance on unseen data and avoid over-fitting.This split was presented as crucial for developing a robust and effective neural network model.
  • 5) Image Slicing: Images and their coco datasets were sliced into smaller components to reduce training time and improve model efficiency.The slicing process used 1000X1000 grids with an overlap ratio of 0.2.
  • 5) Image Slicing: 1000X1000 grids with an overlap ratio of 0.2 were used to divide each image into smaller training components.Figures 6 and 7 showed an original image and a sliced component, while Figure 8 showed the sliced coco dataset.
  • B. ARCHITECTURE AND NEURAL NETWORK MODEL SELECTION: Mask R-CNN was selected as the base algorithm because of its high object-detection accuracy and ability to detect and segment image objects.The study identified architecture and model selection as critical to developing an effective intelligent MEP-metric detection system.
  • B. ARCHITECTURE AND NEURAL NETWORK MODEL SELECTION: Mask R-CNN creates masks outlining the exact pixels belonging to objects, distinguishing it from Faster R-CNN, YOLO, and SSD.It is described as a deep-learning algorithm for both object detection and image segmentation.
  • B. ARCHITECTURE AND NEURAL NETWORK MODEL SELECTION: Mask R-CNN performs object detection and instance segmentation in one model, provides higher accuracy and more detailed object information, but is computationally expensive and slower.Its slower performance is attributed to its two-stage architecture and more complex computations.
  • B. ARCHITECTURE AND NEURAL NETWORK MODEL SELECTION: Mask R-CNN was judged appropriate for this project because its accuracy and segmentation capabilities outweighed its slower, computationally expensive operation.The passage frames model selection as a trade-off between accuracy and computational cost based on application requirements.

C. NEURAL NETWORK MODEL DEVELOPMENT … 3) Compute Loss

The neural network model development phase trained Mask R-CNN to detect and segment MEP metrics through initialization, forward propagation, and loss computation. Training outputs and loss components were logged to monitor progress and adjust model parameters.

  • C. NEURAL NETWORK MODEL DEVELOPMENT: The model development phase trained the Mask R-CNN algorithm to detect and segment MEP metrics through four stages.The supplied passage identifies this phase and states that it involved four stages.
  • 1) Initialization: Weights and biases were randomly initialized to small values to help avoid getting stuck in local optima.This initialization step was part of the model development process.
  • 2) Forward Propagation: During forward propagation, the network processed input images through the network to produce predictions.The predictions were generated by performing a forward pass through the network.
  • 2) Forward Propagation: The forward-pass outputs included predicted bounding boxes, objectness scores, and masks for each object in the input image.These outputs correspond to the object detection and segmentation tasks described in the passage.
  • 3) Compute Loss: A loss function measured differences between predicted and true outputs, with a loss value calculated and printed for each batch during every training epoch.The loss values and other training information were stored in log files.
  • 3) Compute Loss: The logs recorded region proposal network classification loss, bounding box regression loss, and classification accuracy to monitor training and adjust model parameters.These components provided information on different parts of the object detection task.

4) Backpropagation · D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING

The model used stochastic gradient descent with momentum to backpropagate loss and update network parameters. Validation-based optimization tuned hyperparameters using AP and AR, after which the optimized model was independently tested for lighting-symbol detection and generalization.

  • 4) Backpropagation: Loss gradients were computed with respect to each neuron’s weights and biases during backpropagation.The backpropagation phase calculated gradients of the loss for network weights and biases.
  • 4) Backpropagation: Stochastic gradient descent with momentum updated the network’s weights and biases in the loss-minimizing direction.The algorithm iteratively adjusted parameters based on gradients calculated during backpropagation.
  • D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING: The validation dataset evaluated model accuracy and supported tuning of the learning rate, evaluation interval, and number of epochs.Hyperparameter optimization aimed to improve model performance.
  • D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING: Hyperparameters were tuned by trial and error until the desired performance was reached according to Average Precision (AP) and Average Recall (AR).The tuning process sought the combination of hyperparameters producing optimal performance.
  • D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING: AP measured average precision across IoU thresholds, while AR measured average recall across IoU thresholds during optimization.Both metrics were used to assess object-detection performance.
  • D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING: The optimized model was tested on a separate test dataset to assess lighting-symbol detection and performance generalization to new data.This independent evaluation was intended to provide a robust assessment after hyperparameter tuning.
  • D. MODEL EVALUATION, OPTIMIZATION, VALIDATION AND HYPERPARAMETER TUNING: The final optimized model accurately classified lighting symbols and showed high-confidence applicability to real-world MEP detection scenarios.The reported result was described as excellent prediction ability.

IV. RESULTS AND ANALYSIS · A. ANALYSIS OF MODEL PERFORMANCE DURING TRAINING · B. PERFORMANCE EVALUATION

The results show that the Mask R-CNN-based model learned effectively during training, with decreasing losses and improving accuracy, and achieved strong detection and segmentation performance for MEP metrics. Evaluation across annotation types and object sizes further demonstrates its effectiveness.

  • IV. RESULTS AND ANALYSIS: The results section presents findings, analyzes patterns and trends, and discusses implications related to the research objectives.It aims to provide a clear presentation of the results and their significance.
  • A. ANALYSIS OF MODEL PERFORMANCE DURING TRAINING: Training analysis covered loss values, accuracy, and memory usage across 20 epochs to assess model behavior.TensorBoard was used to monitor and visualize training performance interactively.
  • A. ANALYSIS OF MODEL PERFORMANCE DURING TRAINING: The learning-rate graph depicted changes over time and helped determine an appropriate learning rate for optimal performance.The graph was generated by TensorBoard.
  • A. ANALYSIS OF MODEL PERFORMANCE DURING TRAINING: The training loss decreased steadily over time, indicating that the model learned from the data and improved its performance.Figure 10(a) presents accuracy, while Figures 10(b) to (g) show training-loss graphs generated by TensorBoard.
  • A. ANALYSIS OF MODEL PERFORMANCE DURING TRAINING: loss_rpn_cls, loss_rpn_bbox, loss_cls, and loss_bbox decreased significantly with each epoch, while accuracy improved.loss_mask decreased slightly during the first epoch and then remained relatively stable.
  • B. PERFORMANCE EVALUATION: 0.760 mAP and 0.810 mAR were achieved for bounding box annotations at an IoU threshold of 0.50:0.95 with a maximum detection of 100.These metrics were used to evaluate the proposed intelligent detection model for MEP metrics.
  • B. PERFORMANCE EVALUATION: 0.711 mAP and 0.765 mAR were achieved for segmentation annotations at the same IoU threshold and maximum detection.The evaluation used mean Average Precision and mean Average Recall.
  • B. PERFORMANCE EVALUATION: For small, medium-sized, and large objects, mAP values were 0.700, 0.662, and 0.780, while mAR values were 0.700, 0.725, and 0.830, respectively.The results demonstrate the proposed model’s effectiveness in detecting MEP metrics across object sizes.

C. EVALUATION RESULTS · D. DETECTION AND EXTRACTION OF LIGHTING SYMBOLS FOR ENERGY CONSUMPTION PREDICTION

The model was evaluated with detection and segmentation metrics and achieved strong performance overall and at specified IoU thresholds. On an unseen floor-plan image, it identified lighting symbols and extracted associated text for light-type information and power-consumption prediction.

  • C. EVALUATION RESULTS: The evaluation analyzed MEP-metric detection using bbox_mAP, segm_mAP, and variations at different IoU thresholds.Evaluation was conducted using the MMDetection framework, with results shown in Figure 12.
  • C. EVALUATION RESULTS: 0.7596 bbox_mAP and 0.7111 segm_mAP demonstrate high model accuracy for bounding-box detection and segmentation.bbox_mAP measures mean average precision for bounding-box detection, while segm_mAP measures mean average precision for segmentation.
  • C. EVALUATION RESULTS: 0.9850 bbox_mAP_50 and 0.9219 segm_mAP_75 show strong performance at different IoU thresholds.bbox_mAP_50 and segm_mAP_75 measure average precision at specified intersection over union thresholds.
  • C. EVALUATION RESULTS: The evaluation results provide insight into accurate MEP-metric detection and identify areas for further improvement.The reported metrics characterize detection and segmentation performance across standard and threshold-specific measures.
  • D. DETECTION AND EXTRACTION OF LIGHTING SYMBOLS FOR ENERGY CONSUMPTION PREDICTION: The trained model successfully identified lighting symbols in an unseen floor-plan image.The unseen-image evaluation output is presented in Figure 13.
  • D. DETECTION AND EXTRACTION OF LIGHTING SYMBOLS FOR ENERGY CONSUMPTION PREDICTION: The identified lighting symbols were extracted, and Pytesseract extracted the text associated with each symbol.The text-extraction result is shown in Figure 14.
  • D. DETECTION AND EXTRACTION OF LIGHTING SYMBOLS FOR ENERGY CONSUMPTION PREDICTION: The extracted text indicates light type and can support prediction of overall power consumption.Successful text extraction demonstrates the model’s effectiveness in identifying and extracting MEP metrics from floor-plan images.

V. CONCLUSION

The study applied Mask RCNN to detect MEP symbols from floor plans, achieving strong detection and segmentation results at multiple IoU thresholds. The approach automates MEP symbol detection, improves workflow efficiency, and establishes groundwork for broader future applications.

  • Method and contribution: Mask RCNN was applied to intelligently detect MEP symbols from floor plans to improve engineering design efficiency, accuracy, and cost-effectiveness.The method targets automated MEP symbol detection within floor-plan images.
  • Performance: 0.7596 bbox_mAP and 0.7111 segm_mAP were achieved, respectively.These were the study’s overall bounding-box and segmentation mean average precision results.
  • Performance: 0.9850 bbox_mAP_50 and 0.9219 segm_mAP_75 were obtained at different IoU thresholds, respectively.The results indicate strong performance under the reported threshold-specific evaluations.
  • Practical impact: The approach automates manual MEP symbol detection, reducing design time and increasing workflow efficiency.The authors identify architecture, construction, and real estate as fields with potential impact.
  • Energy efficiency: The design-process benefits may contribute to the Government of Canada’s goal of net-zero greenhouse gas emissions by 2050.This motivation is stated in the context of improving energy efficiency during design.
  • Future directions: The method represents a substantial advancement in MEP engineering and provides groundwork for improving precision and versatility across image types and qualities.The conclusion frames these improvements as objectives for future studies.
Loading 2608.14317v1…