Method and apparatus for providing realistic visualization of an architectural structure

Abstract

Method and apparatus for providing realistic visualization of an architectural structure is disclosed. The method comprises receiving one or more two- dimensional images indicating floor plan of the architectural structure, pre- processing the received one or more images using a deep learning model, identifying one or more markers from the one or more received images based on an output from the deep learning model, generating three-dimensional (3D) images from the two-dimensional images of the floor plans, identifying a type of room of the architectural structure in the 3D images based on the identified one or more markers, generating a visualization of one or more components associated with the architectural structure in the generated 3D images, generating a walk-through of different rooms of the building in the 3D images using virtual reality techniques, and auto scaling the one or more components in the 3D images based on the pre- defined real-world dimensions.

Description

[0001] METHOD AND APPARATUS FOR PROVIDING REALISTIC VISUALIZATION OF AN ARCHITECTURAL STRUCTURE TECHNICAL FIELD

[0002] The present invention relates to the field of designing floor plans for buildings, and more particularly relates to providing realistic visualization of an architecture structure using deep learning techniques.

[0003] BACKGROUND

[0004]

[0001] In the real-estate industry, analyzing architectural plans plays an important role. Architectural plans comprises, but not limited to, floor plans, tower plans, elevations, society master plans and serves as important piece of information for the users (both purchaser and the seller). The floor plans disclose crucial set of information required for the architectural structure of the building, including the layout of the floor / building, position of the doors, windows, etc. An architect or a user desiring to purchase a house generally looks at the floor plans to identify the structural details about the floor, such as dimensions of the floor, rooms, position of doors, windows, etc.

[0005]

[0002] Since the user desiring to purchase the house solely relies on the floor plans to make the purchase decision, accuracy of the floor plans becomes really important. Further, two dimensional (2D) floorplans generally are not able to give an exact idea about the actual look and feel of the house. This makes it difficult for the user to imagine the house after it has been completed. Furthermore, with 2D floorplans, it becomes difficult for the house builder to explain to the user the final construction layout of the house.

[0006]

[0003] Traditional methods for converting 2D floorplans into three-dimensional (3D) floor plans heavily rely on manual input and intervention. This process is not only time-consuming but also prone to human error, requiring extensive effort fromskilled professionals to ensure accuracy and detail in the final models. Prior systems often struggle with the precise detection and segmentation of critical architectural elements such as walls, doors, and windows. This lack of precision can result in 3D models that do not accurately reflect the true layout and specifications of the physical space, leading to potential misunderstandings in planning and development phases. The scalability of traditional 2D to 3D conversion technologies is often hampered by their reliance on manual processes and the detailed supervision required. As projects increase in size and complexity, these methods become increasingly inefficient, making them less suitable for larger or more complex architectural endeavors.

[0007]

[0004] Prior systems generally exhibit limited flexibility in terms of adapting to varied architectural styles or non-standard design elements. This rigidity can hinder the creation of customized or innovative design solutions, thereby stifling creativity and adaptation in architectural and interior design processes. Similarly, as explained earlier, 3D models of real-estates also benefits a builder or sales person associated with the real-estate to demonstrate exact aesthetics of the real-estate to the users purchasing the same. This might help the user make purchase decisions easily and efficiently. Therefore, the need arises to provide improved techniques to enhance user experience by providing ways by which the user can have an actual look and feel of the house even before the user decides to purchase the house.

[0008] SUMMARY

[0009]

[0005] The following presents a simplified summary of the subject matter in order to provide a basic understanding of some aspects of subject matter embodiments. This summary is not an extensive overview of the subject matter. It is not intended to identify key / critical elements of the embodiments or to delineate the scope of the subject matter.

[0006] Its sole purpose is to present some concepts of the subject matter in a simplified form as a prelude to the more detailed description that is presented later.

[0010]

[0007] In one embodiment, a method for providing realistic visualization of an architectural structure is disclosed. The method comprises receiving one or more two-dimensional images indicating floor plan of the architectural structure, preprocessing the received one or more images using a deep learning model, identifying one or more markers from the one or more received images based on an output from the deep learning model, generating three-dimensional (3D) images from the two-dimensional images of the floor plans, identifying a type of room of the architectural structure in the 3D images based on the identified one or more markers, generating a visualization of one or more components associated with the architectural structure in the generated 3D images, generating a walk-through of different rooms of the building in the 3D images using virtual reality techniques, and auto scaling the one or more components in the 3D images based on the predefined real-world dimensions.

[0011]

[0008] In another embodiment, an apparatus for providing realistic visualization of an architectural structure is provided. The apparatus comprises a receiving module configured to receive one or more images indicating floor plan of the architectural structure, wherein the one or more images are two-dimensional images of floor plan, a pre-processing module is configured to pre-process the received one or more images using a deep learning model, a memory and a processor. The processor is configured to identify one or more markers from the one or more received images based on an output from the deep learning model, generate three-dimensional (3D) images from the two-dimensional images of the floor plans, identify a type of room of the architectural structure in the 3D images based on the identified one or more markers, generate a visualization of one or more components associated with the architectural structure in the generated 3D images, generate a walk-through of different rooms of the building in the 3D images using virtual reality techniques,and auto scale the one or more components in the 3D images based on the predefined real-world dimensions.

[0012]

[0009] These and other objects, embodiments and advantages of the present invention will become readily apparent to those skilled in the art from the following detailed description of the embodiments having reference to the attached figures, the invention not being limited to any particular embodiments disclosed.

[0013] BRIEF DESCRIPTION OF FIGURES

[0014]

[0010] The foregoing and further objects, features and advantages of the present subject matter will become apparent from the following description of exemplary embodiments with reference to the accompanying drawings, wherein like numerals are used to represent like elements.

[0015] [OH] It is to be noted, however, that the appended drawings along with the reference numerals illustrate only typical embodiments of the present subject matter, and are therefore, not to be considered for limiting of its scope, for the subject matter may admit to other equally effective embodiments.

[0016]

[0012] Figure 1 illustrates a system for providing realistic visualization of an architectural structure, according to an embodiment of the present invention.

[0017]

[0013] Figure 2 illustrates a flowchart of an implementation pipeline, according to an embodiment of the present invention.

[0018]

[0014] Figure 3 illustrates an output from deep learning model, in accordance with one embodiment of the present invention.

[0015] Figures 4(a)-3(e) illustrates step by step process of the computer vision postprocessing, according to an embodiment of the present invention.

[0019]

[0016] Figure 5 illustrates the creation of final 3D rendering, according to an embodiment of the present invention.

[0020]

[0017] Figure 6 illustrates a network architecture of nnuNet, according to an embodiment of the present invention.

[0021]

[0018] Figure 6 illustrates a flowchart of a method for converting two-dimensional floor plans into three-dimensional model, according to an embodiment of the present invention.

[0022]

[0019] Figure 7 illustrates a flowchart of a method for providing realistic visualization of an architectural structure, according to an embodiment of the present invention.

[0023]

[0020] Figure 8 illustrates a block diagram of a computing device, according to an embodiment of the present invention.

[0024] DETAILED DESCRIPTION

[0025]

[0021] Exemplary embodiments now will be described with reference to the accompanying drawings. The disclosure may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey its scope to those skilled in the art. The terminology used in the detailed description of the particular exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting. In the drawings, like numbers refer to like elements.

[0022] It is to be noted, however, that the reference numerals used herein illustrate only typical embodiments of the present subject matter, and are therefore, not to be considered for limiting of its scope, for the subject matter may admit to other equally effective embodiments.

[0026]

[0023] The specification may refer to “an”, “one” or “some” embodiment s) in several locations. This does not necessarily imply that each such reference is to the same embodiment(s), or that the feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments.

[0027]

[0024] As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless expressly stated otherwise. It will be further understood that the terms “includes”, “comprises”, “including” and / or “comprising” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0028]

[0025] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0029]

[0026] The figures depict a simplified structure only showing some elements and functional entities, all being logical units whose implementation may differ from what is shown. The connections shown are logical connections; the actual physicalconnections may be different. It is apparent to a person skilled in the art that the structure may also comprise other functions and structures.

[0030]

[0027] The present invention provides for techniques for converting two-dimensional (2D) floor plans into a three-dimensional (3D) model and identifying the wall / window / door, etc. from the 2D plan based on the deep neural network segmentation algorithms. The present invention further provides techniques for identifying and classifying different rooms from the floorplan represented in the 3D model using graph neural network algorithms, provide visualization of furniture, wall, floor ceiling in the 3D model. The present invention further provides techniques for enabling walk-through of the 3D model using VR technology provider visualization of other features of the property, unit plan basis and apartment level basis, such as landscape views of the society, elevation views. The present invention further provides techniques for autoscaling furniture automatically detected from image information and using reference structures to scale every other asset to scale.

[0031]

[0028] Referring to Figure 1, an apparatus 100 for providing realistic visualization of an architectural structure in accordance with one embodiment of the present invention is disclosed. The system 100 comprises a receiving module 102, a preprocessing module 104, a memory 106 and a processor 108. In one embodiment, the processor 108 is coupled to the receiving module 102, the pre-processing module 104 and the memory 106. The components in the system 100 are not limited to the one mentioned here and may involve other components as well. In one embodiment, the system 100 is a computer system, a smartphone, an iPad, a tablet device, a laptop, etc.

[0032]

[0029] The receiving module 102 is configured to receive one or more images indicating floor plan of the architectural structure, where the one or more images are two-dimensional images of floor plan. The one or more images may begenerated by an architect using line drawings or using one or more software. The floor plan includes layout of the floor including the demarcations of different rooms. The demarcations are shown by one or more markers. Further, other details such as dimensions, etc. are also mentioned in the floor plans.

[0033]

[0030] The pre-processing module 104 is configured to pre-process the received one or more images using a deep learning model. The present invention combines image processing techniques with advanced deep neural network algorithms to precisely analyze and segment 2D spatial images. The pre-processing may involve one or more techniques such as noise reduction, image resizing, contrast enhancement, edge detection, etc. The one or more images of the floor plan are fed to the pre-processing module 104 for performing pre-processing using any of the techniques as described above.

[0034]

[0031] This integration facilitates the accurate identification of critical structural elements such as walls, doors, and windows, ensuring a high-fidelity base for subsequent modelling stages. The deep learning model orchestrates a sophisticated suite of algorithms to process segmented data effectively. This includes the vectorization of the floorplan into detailed segmentation maps and the transformation of these maps into dynamic 3D models. This orchestrated process enhances both the accuracy and the quality of the final 3D representations.

[0035]

[0032] The pre-processing also includes collecting images of the floor plan and removing disturbances from the images. Once the disturbance is removed, the preprocessing includes labelling floorplan data pixel wise for walls, doors and windows. The deep learning model may be trained for performing pixel wise semantic segmentation and room classification.

[0033] The processor is configured to identify one or more markers from the one or more received images based on an output from the deep learning model. The one or more markers may already be provided in the two-dimensional drawings of the floor plans. The one or more markers provide details about the boundaries of various rooms present in the floor plan, the presence and location of doors, windows, etc. Once the deep learning model performs pre-processing by applying the one or more techniques as explained above, the output of the deep learning model can be used to identify one or more markers from the one or more images.

[0036]

[0034] Once the markers are identified, the processor is configured to generate three-dimensional (3D) images from the two-dimensional images of the floor plans. The types of rooms are identified in the 3D images based on the one or more identified markers. The type of room may include living room, bedroom, dining room, etc. In some cases, the name of the rooms may be mentioned in the floor plans. In one embodiment, the name of the rooms may be mentioned in abbreviations format. In another embodiment, the type of room may be identified based on size of the room. In this case, the deep learning model may be trained to identify type of room based on the size of the room. For example, the room with biggest size may be identified as living room, the area near to the kitchen may be identified as dining room / lobby area, etc.

[0037]

[0035] The present invention provides a novel enhancement to the existing floor plan analysis pipeline through the integration of Vision-Language Models (VLMs) as a verification and correction mechanism. This approach addresses a critical limitation in current deep learning segmentation methods, which can fail to properly identify structural elements when pixel-wise segmentation produces incomplete or inaccurate results. By incorporating VLMs at strategic points in the processing pipeline, the system achieves significantly improved robustness and accuracy in wall detection, room identification, and overall floor plan interpretation.

[0036] The VLM reinforcement operates through a targeted inspection framework rather than attempting complete segmentation generation. When the primary segmentation model produces a graph representation with potential errors or omissions, the system overlays this preliminary graph onto the original floor plan image and presents it to the VLM component. The VLM then performs specific verification tasks: identifying missing walls, confirming room boundaries, and validating architectural coherence. For each identified discrepancy, the system applies enhanced segmentation sensitivity specifically in the regions of interest, effectively creating a feedback loop that iteratively improves the structural representation. This approach leverages the VLM's semantic understanding of architectural conventions and spatial relationships without requiring it to generate pixel-perfect segmentation maps, which would be ill-suited to its capabilities. The result is a hybrid system that combines the precision of specialized segmentation algorithms with the contextual awareness of large vision-language models, significantly reducing the occurrence of incomplete room boundaries and structural inconsistencies in the final 3D model.

[0038]

[0037] Once the markers and the type of rooms have been identified, the visualization of one or more components associated with the architectural structure in the generated 3D images is generated. In one embodiment, the one or more markers are identified using mnUNet architecture.

[0039]

[0038] The visualization of one or more components may include visualization of furniture, walls, floor ceilings in the 3D model so that the user can have the real feel of the structure of the floor plan. The 3D visualization of the floor plan provides details about the actual structure of the building / floor.

[0040]

[0039] Once the 3D visualization has been generated, the processor is configured to generate walk-through of different rooms of the building using virtual reality technique. For experiencing a virtual reality-based walk-through, a user may usevirtual reality (VR) based glasses. In one embodiment, the user may wear the VR glasses which would enable the user to experience the floor as if the user is present on the floor. In one embodiment, the user may be represented by a virtual avatar in a VR environment. The user may use controls present with the user to navigate in the virtual environment. In one embodiment, the user may view the virtual environment on a display screen and may use the controls present on the display screen to navigate through the virtual environment. With the help of virtual reality techniques, the user is able to virtually visit the actual building / floor without physically visiting the place.

[0041]

[0040] The processor is also configured to perform auto scaling the one or more components in the 3D images based on the pre-defined real-world dimensions. This would involve virtually adjusting the furniture present in the building / floor so that the user can experience the actual placement of the furniture in the building / floor. This would provide the user with better experience of how the placement of the furniture would look. For this, the user may be provided with features on the display screen to place the different representation of the furniture from furniture library. The autoscaling is performed by performing pixel-wise measurement of the one or more markers present in the 2D image, extracting one or more dimensions of the one or more markers present in the 2D image, determining a scaling factor based on the pixel-wise measurement and the extracted one or more dimensions, and performing auto-scaling of the one or more components based on the scaling factor.

[0042]

[0041] The present invention introduces an intuitive real-time furniture placement system that enhances the 3D visualization experience. Users can access a comprehensive furniture library containing various categories of pre-modeled furnishings and decorative elements. The system enables seamless drag-and-drop functionality directly within the 3D environment, allowing users to select items from the library and place them precisely within the generated space. This feature helps introduce contextual awareness that automatically detects appropriateplacement zones (e.g., walls for cabinets, floor areas for seating), Real-time collision detection preventing object interpenetration while maintaining realistic positioning, intelligent snapping functionality that aligns furniture to walls, corners, and other architectural elements. This feature also helps with on-the-fly rotation, scaling, and elevation controls accessible through intuitive gestures, automatic arrangement suggestions based on room type identification and spatial optimization, algorithms Material and finish customization options for each furniture piece to match desired aesthetic themes and dynamic lighting adjustment showing how furniture interacts with natural and artificial light sources in the space.

[0043]

[0042] Referring now to Figure 2, a flowchart of an implementation pipeline 200 according to an embodiment of the present invention is shown. At step 202, the implementation pipeline 200 comprises receiving input images. The input images includes two-dimensional images of the architectural drawings including the floor plans. The input images may be in JPG, JPEG or PNG formats.

[0044]

[0043] At step 204, the images are fed to the deep learning model for further processing. The 2-dimensional images are considered as data sets for feeding into the deep learning model. Traditional methods typically require large and comprehensive datasets for training the conventional algorithms, which are not always readily available or accessible. This dependence can limit the applicability of such systems in regions or scenarios where such datasets are scarce, ultimately restricting the versatility and deployment of these technologies.

[0045]

[0044] Further, at step 206, the computer vision pre-processing is performed. Since the steps 204 and 206 are interconnected, they have been explained together. The present invention combines image processing techniques with advanced deep neural network algorithms to precisely analyze and segment 2D spatial images. This integration facilitates the accurate identification of critical structural elements such as walls, doors, and windows, ensuring a high-fidelity base for subsequent modelling stages. The deep learning model orchestrates a sophisticated suite ofalgorithms to process segmented data effectively. This includes the vectorization of the floorplan into detailed segmentation maps and the transformation of these maps into dynamic 3D models. This orchestrated process enhances both the accuracy and the quality of the final 3D representations.

[0046]

[0045] The deep learning model and the computer vision postprocessing techniques help to raster images of floor plans into precise vector data. This includes advanced techniques such as single class stacking, largest connected component extraction, and skeletonization to produce a detailed segmentation map. The deep learning model features an advanced Wall / Window / Door Pixel wise Segmentation Model that accurately identifies these key structural elements within the 2D plan. The model has been trained on 5000 samples with adversarial examples where the segmentation model could perform poorly - this involves finding the kind of floorplans the segmentation model is performing poorly on and training the model with higher weightage on those images specifically. The architecture of the segmentation model is based on a U-Net class of models which are popularly used for this class of problems. Coupled with a Room Classification Model, the present invention helps in efficiently categorizing different areas of a floor plan, enhancing the semantic richness of the 3D model.

[0047]

[0046] The robust computer vision post-processing steps helps to refine the segmentation map, including corner extraction using the Harris Corner Detection algorithm, line detection, and merging of redundant lines. These steps ensure precise and clean fitting of structural components, resulting in highly accurate 3D models. After initial model generation, the present invention helps to conduct mesh clean-up and model cutting to remove any artifacts and ensure geometric accuracy. Rigging of lighting and reflections is also performed, which significantly enhances the realism and visual appeal of the rendered models.

[0047] The present invention includes a comprehensive material library and a collection of design assets. This enables users to customize surfaces and interiors with high-quality textures and finishes, further enhancing the aesthetic quality and realism of the 3D models. Utilizing the nnUNet architecture for semantic segmentation, Depth-Craft achieves exceptional accuracy in detecting and delineating various elements within the floorplan. The system has demonstrated a Dice score of 90.56% on an independent test set, signifying a substantial enhancement in segmentation precision compared to existing technologies. The selection of nnUNet architecture for this invention was driven by several key factors. First, nnUNet’ s ability to automatically configure and optimize hyperparameters based on dataset properties eliminates the need for manual tuning, making the system more robust across varying floor plan styles and formats. Second, the architecture’s dynamic adaptation capability ensures good performance regardless of input image resolution or quality differences commonly found in architectural drawings. Finally, nnUNet uses an ensemble learning approach, combining multiple network variations and thus provides superior segmentation results compared to single-model approaches.

[0048]

[0048] Referring to FIG. 3, output from the deep learning model (step 204) is explained in detail. As shown, the 3 class segmentation is highlighted by different colors. The output from the deep learning model is used to create 3D rendering information. The model output is tested on independent training set returned performance score of 90.56% Dice Score which is a metric for quality of segmentation.

[0049]

[0049] Referring to Figs 4(a)-4(e), a step-by-step process of the computer vision postprocessing (step 206) is shown. FIG. 4(a) shows the step of single class stacking output where structural information is extracted from deep learning model output through stacking temporarily into single channel to extract wall indices. FIG. 4(b) shows the second step where disturbances are removed from stacked image to extract largest connected component. In FIG. 4(c), skeleton mapping of image iscreated through thinning lines by iterative reduction with multiple passes through single image.

[0050]

[0050] Next in FIG. 4(d), corner extraction is performed where Harris comer detection algorithm is used to create structural elements after which they are axis aligned to nearest axis for detecting corners. Next in FIG. 4(e), the points are iterated on in two scans to create connections based on point location to detect if there are connections in input image. After detecting possible lines, redundant lines are merged to create unique walls. Similar procedure is done for windows and doors.

[0051]

[0051] Referring back to FIG. 2, at step 208, the implementation pipeline comprises providing 3D rendering based on the output from the computer vision post processing. Integrating with rendering libraries like Three.js, the present invention provides an immersive interface that allows users to interactively explore the 3D model. This feature supports functionalities akin to a 3D game, including walking through the model or using orbit controls to achieve an isometric view, significantly enriching the user experience. FIG. 5 shows the creation of the final 3D rendering based on the output from the computer vision post processing.

[0052]

[0052] The present invention implements a novel approach to furniture autoscaling that allows precise dimensional accuracy in a 3D setting scaled to real world dimensions. This automated scaling system operates by first establishing dimensional benchmarks through analyzing standardized architectural elements within the 2D floorplan. The system performs pixel-wise measurement of structural elements such as doors, wall thicknesses, and other standard architectural features. Additionally the system employs Optical Character Recognition (OCR) technology to extract any explicitly stated dimensions from the floor plan text annotations. The scaling process then uses the known real -world dimensions of standard architectural elements - particularly doorways, which typically measure between 0.75 to 0.9 meters in width - as reference points. Using these standardized elements as baselinemeasurements, the system calculates a precise scaling factor by comparing the pixel measurements of these elements to their known real-world dimensions. This calculated scaling factor is then uniformly applied across the entire floor plan to ensure dimensional consistency. This implementation allows the system to maintain a precise proportional relationship throughout the 3D model. Wall heights are also rendered according to known real world standard dimensions and pre-scaled 3D furniture models from the asset library are automatically adjusted to match these real -world dimensions. This systematic approach to scaling makes it such that when furniture models are placed in the 3D environment, they naturally align with the architectural dimensions of the space. This automated dimensional calibration is particularly valuable when performing virtual walkthroughs to provide an accurate sense of spatial relationships and room proportions.

[0053]

[0053] The preset invention advances beyond current technologies by seamlessly integrating multiple advanced technologies — image processing, deep learning, and interactive 3D rendering — into a unified platform that transforms 2D architectural drawings into highly accurate and interactive 3D models. The preset invention not only enhances the precision and efficiency of architectural modelling but also democratizes complex 3D visualization, making it accessible and practical for a broader range of users in the real estate and architectural industries.

[0054]

[0054] Additionally, the present invention enables highly immersive VR walkthroughs of architectural designs, transforming how users experience and interact with spaces. This capability is especially beneficial for real estate presentations, architectural reviews, and providing virtual tours that offer a realistic sense of space and layout without physical site visits.

[0055]

[0055] The invention significantly enhances real estate listings by providing interactive 3D models of properties. These models allow potential buyers to virtually explore properties, offering a comprehensive understanding of theproperty's layout, design, and potential, thereby facilitating informed purchasing decisions.

[0056]

[0056] The present invention assists architects in visualizing and planning building spaces by converting traditional 2D floorplans into detailed 3D models. This tool enables architects to make better-informed design decisions, optimize space usage, and improve client presentations by showcasing lifelike representations of architectural projects.

[0057]

[0057] The preset invention aids interior designers in creating, visualizing, and modifying room layouts in a dynamic 3D environment. Depth-Craft's capability to apply various textures and materials, and to view designs from multiple angles, enhances the design process, allowing for immediate revisions and experimentation with different styles.

[0058]

[0058] By offering 3D visualizations, the preset invention helps professionals in real estate and design industries engage clients more effectively, providing them with a tangible understanding of properties and design projects. The invention streamlines the workflow of designing and revising architectural and interior projects, reducing time and costs associated with physical model creation and iterative changes based on client feedback. By automating the conversion of 2D drawings to 3D models, the invention minimizes human error and increases the accuracy of architectural and design representations, ensuring that the final constructions closely align with their intended designs.

[0059]

[0059] Referring to FIG. 6, network architecture 600 of the nnUNet is shown. This architecture represents the core of the deep learning model employed by the present invention, which is meticulously trained on labeled floorplan images. The figure illustrates how the nnUNet segments architectural elements such as walls, doors,and windows in a pixel-wise manner, highlighting the sophisticated computational processes that underpin the functionality of the system.

[0060]

[0060] At block 602, the input data is fed into deep learning model. This process starts by feeding the data (2D / 3D medical images or architectural floorplans) into the deep learning model. At block 604, the input data is trained by the deep learning model. The deep learning model uses data Fingerprinting techniques (Heuristic Rules). The nnUNet applies heuristic rules to perform a data fingerprinting step, analyzing the dataset's properties (e.g., resolution, modality) to generate metadata describing the data. This metadata informs how the network configuration should be adjusted for the specific dataset.

[0061]

[0061] The images may be resampled to ensure consistent resolution across training samples. Standard normalization techniques are applied to improve training efficiency. Depending on the dataset's complexity, the network architecture is adapted dynamically. These parameters are adjusted according to the dataset's properties to optimize memory and computation. The error measurement during training are also specified. The optimizer is chosen to minimize the loss function and train the network efficiently. A specific network design is used as a starting template.

[0062]

[0062] These parameters determine how the network will be trained over time and how data augmentation will enhance generalization. Each pipeline configuration (specific to the dataset) is tracked as a "fingerprint" that records which parameters were used in training. Multiple networks (2D, 3D, and 3D cascaded networks) are trained on the data using cross-validation. This ensures robust model performance across different subsets of the data.

[0063]

[0063] At blocks 606 and 608, postprocessing is performed on the input data. Additional operations are applied to improve the segmentation output (e.g., removing small segmentation artifacts). Multiple models are combined to create afinal prediction, leveraging the strengths of different architectures. After the model is trained and optimized, it can predict on new test data, applying the learned patterns from the training process.

[0064]

[0064] Referring to FIG. 7, a flowchart of a method 700 for converting two-dimensional floor plans into three-dimensional model is disclosed. At step 702, the method comprises collecting two-dimensional (2D) floor plan images and removing disturbances from the collected floor plan images. The 2D floor plans may be collected by architects, builders, and other uses linked with the real-estates. The disturbances may be removed by performing pre-processing on the collected floor plan images.

[0065]

[0065] At step 704, the method comprises labelling floorplan data pixel wise for walls, doors and windows. At step 706, the pixel wise semantic segmentation is performed to train a convolution neural network (CNN) model. With pixel wise semantic segmentation, each pixel is assigned a label or a class so that similar pixels can be grouped together.

[0066]

[0066] At step 708, the method comprises training room classification graph neural network. This classification model is trained to classify different rooms that are present in the 2D floorplans. As discussed above, the floor plans has different label for the rooms. Form the different labels provided for the rooms, CNN model can be trained to classify different rooms.

[0067]

[0067] At step 710, the method comprises performing objective metrics calculation based on the output from steps 706 and 708. This step is using a objective metric of quality of pixel wise segmentation model from 708 using a metric called Dice Score which is calculated by comparing output of the model to the ground truth i.e. perfect output and the quality of the room classification model i.e. 708 is determined using accuracy which is the number of rooms that are correctly classified. At step 712, the method comprises performing computer vision post-processing pipeline to cleanup output of CNN model which is an orchestration of multiple image cleaning operations to extract structural information from pixel level segmentation. At step 714, the method comprises rendering three-dimensional (3D) model in ThreeJS with output of deep learning models. At step 716, the method comprises creating automatic furniture templates with room type heuristics, this involves creating meaningful configuration of assets for each type of room and the amount of space available in relation to space and adjacency of other rooms.

[0068]

[0068] Referring to FIG. 8 now, a method for providing realistic visualization of an architectural structure is provided. At step 802, the method comprises receiving one or more images indicating floor plan of the architectural structure. The one or more images of the floor plan are two-dimensional images of floor plan. At step 804, the method comprises pre-processing the received one or more images using a deep learning model. The pre-processing includes, but not limited to, noise reduction, image resizing, contrast enhancement, edge detection, etc.

[0069]

[0069] At step 806, the method comprises identifying one or more markers from the one or more received images based on an output from the deep learning model. The one or more markers include, but not limited to, wall, window and / or door of the building. In one embodiment, the one or more markers are identified using mnUNet architecture. At step 808, the method comprises generating three-dimensional (3D) images from the two-dimensional images of the floor plans.

[0070]

[0070] At step 810, the method comprises identifying a type of room of the architectural structure in the 3D images based on the identified one or more markers. The type of room may include the living room, dining room, bedroom, etc. At step 812, the method comprises generating a visualization of one or more components associated with the architectural structure in the generated 3D images. Once the visualization has been generated, the method comprises, at step 814, generating a walk-through of different rooms of the building in the 3D images usingvirtual reality techniques. At step 816, the method comprises auto scaling the one or more components in the 3D images based on the pre-defined real-world dimensions. The auto-scaling is performed by performing pixel-wise measurement of the one or more markers present in the 2D image, extracting one or more dimensions of the one or more markers present in the 2D image, determining a scaling factor based on the pixel-wise measurement and the extracted one or more dimensions, and performing auto-scaling of the one or more components based on the scaling factor.

[0071]

[0071] Referring to FIG. 9, a block diagram of an exemplary computing device 900 in which one or more embodiments of the present invention may operate, according to an embodiment. In the system schematic of figure 9, bus 910 is in physical communication with Input / Output device 902, interface 904, memory 906, and processor 908. Bus 910 includes a path that permits components within computing device 900 to communicate with each other. Examples of Input / Output device 902 include peripherals and / or other mechanism that may enable a user to input information to computing device 900, including a keyboard, computer mice, buttons, touch screens, voice recognition, and biometric mechanisms and the like Input / Output device 902 also includes a mechanism that outputs information to the user of computing device 900, such as a display, a light emitting diode (LED), a printer, a speaker, and the like.

[0072]

[0072] Examples of interface 904 include mechanisms that enable computing device 900 to communicate with other computing devices and / or systems through network connections. Examples of memory 906 include random access memory (RAM), read-only memory (ROM), flash memory, and the like. The memory 906 store information and instructions for execution by processor 908. The processor 908 includes, but not limited to, a microprocessor, an application specific integrated circuit (ASIC), or a field programmable object array (FPOA) and the like. The processor 808 interprets and executes instructions retrieved from memory 906.

[0073] In one embodiment, the computing device 900 may be responsible for implementing the above-mentioned steps. For example, input parameters of the user may be received using the Input / Output device 902. The deep learning models may be stored in memory 906 and may be implemented by processor 908.

[0073]

[0074] In present invention, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium and executed by a hardwarebased processor. The implementations described herein are not limited to any specific combinations of hardware circuitry and software.

[0074]

[0075] In the drawings and specification, there have been disclosed exemplary embodiments of the invention. Although specific terms are employed, they are used in a generic and descriptive sense only and not for purposes of limitation of the scope of the invention.

Claims

CLAIMSWE CLAIM1. A method for providing realistic visualization of an architectural structure, the method comprising:receiving one or more images indicating floor plan of the architectural structure, wherein the one or more images are two-dimensional images of floor plan;pre-processing the received one or more images using a deep learning model;identifying one or more markers from the one or more received images based on an output from the deep learning model;generating three-dimensional (3D) images from the two-dimensional images of the floor plans;identifying a type of room of the architectural structure in the 3D images based on the identified one or more markers;generating a visualization of one or more components associated with the architectural structure in the generated 3D images;generating a walk-through of different rooms of the building in the 3D images using virtual reality techniques; andauto scaling the one or more components in the 3D images based on the pre-defined real-world dimensions.

2. The method as claimed in claim 1, wherein the one or more markers include, but not limited to, wall, window and / or door of the building.

3. The method as claimed in claim 1, wherein the one or more components include, but not limited to, furniture, walls, floor ceiling.

4. The method as claimed in claim 1, wherein the one or more markers are identified using mnUNet architecture.

5. The method as claimed in claim 1, wherein the autoscaling is performed by:performing pixel-wise measurement of the one or more markers present in the 2D image,extracting one or more dimensions of the one or more markers present in the 2D image,determining a scaling factor based on the pixel-wise measurement and the extracted one or more dimensions, andperforming auto-scaling of the one or more components based on the scaling factor.

6. The method as claimed in claim 1, further comprises performing verification of the generated 3D images based on vision-language models (VLMs).

7. An apparatus for providing realistic visualization of an architectural structure, the method comprising:a receiving module configured to receive one or more images indicating floor plan of the architectural structure, wherein the one or more images are two-dimensional images of floor plan;a pre-processing module is configured to pre-process the received one or more images using a deep learning model;a memory;a processor configured to:identify one or more markers from the one or more received images based on an output from the deep learning model;generate three-dimensional (3D) images from the two-dimensional images of the floor plans;identify a type of room of the architectural structure in the 3D images based on the identified one or more markers;generate a visualization of one or more components associated with the architectural structure in the generated 3D images;generate a walk-through of different rooms of the building in the 3D images using virtual reality techniques; andauto scale the one or more components in the 3D images based on the pre-defined real-world dimensions.

8. The apparatus as claimed in claim 6, wherein the one or more markers include, but not limited to, wall, window and / or door of the building.

9. The apparatus as claimed in claim 6, wherein the one or more markers are identified using mnUNet architecture.

10. The apparatus as claimed in claim 6, wherein the autoscaling is performed by:performing pixel-wise measurement of the one or more markers present in the 2D image,extracting one or more dimensions of the one or more markers present in the 2D image,determining a scaling factor based on the pixel-wise measurement and the extracted one or more dimensions, andperforming auto-scaling of the one or more components based on the scaling factor.

Images