Computational parsing of architectural floor plans has taken a substantial step forward with the introduction of the Explicit-boundary Dual-stream Attention Network, an artificial intelligence architecture designed to resolve chronic geometric distortions in building layout analysis. Developed to convert floor plan drawings into structured, machine-interpretable spatial maps, the new system directly tackles two primary historical stumbling blocks in automated layout parsing: structural disconnections along extended corridors and the unchecked spilling of room semantic labels across physical wall divisions. Across comprehensive evaluations on standard benchmark datasets, the framework established new performance records, achieving mean Intersection over Union scores of 78.0 percent on the textured Rent3D benchmark and 81.0 percent on the line-based Raster-to-Vector dataset. By outperforming established convolutional frameworks as well as modern Vision Transformer paradigms, the model demonstrates how explicitly integrating directional feature pooling alongside mathematically constrained geometric boundaries allows computer vision algorithms to reliably trace building layouts ranging from rich real estate floor plans to sparse monochrome engineering blueprints.
In contemporary architectural practice, property management, and automated building information modeling, accurate floor plan interpretation serves as an essential foundation for robotic spatial navigation and environmental perception. For autonomous systems operating indoors, automated models must reliably reconstruct functional navigation maps, distinguishing private bedrooms, dining areas, and kitchens while isolating structural boundaries such as partition walls, doors, and windows. Historically, extracting these spatial representations required painstaking manual tracing and visual inspection, an operational bottleneck characterized by low efficiency, high labor overhead, and subjective discrepancies among human annotators. While computer vision methods have increasingly automated this workflow, practical deployment continues to confront drastic variations in architectural drawing styles. Floor plans typically present either as visually cluttered illustrations loaded with decorative furniture symbols, material textures, and text annotations that obscure physical perimeters, or conversely, as minimalist line drawings stripped of semantic surface cues, forcing algorithms to infer room function strictly from sparse geometric outlines.
Earlier computational efforts to process architectural drawings relied on classical image processing pipelines, employing heuristic edge detectors like the Canny operator, morphological operations, and Hough transforms to isolate geometric primitives. While functional for standardized vector drawings, these hand-crafted rules collapsed when confronted with stylistic diversity, scanned drawing noise, and complex spatial constraints, such as doors invariably embedding within structural walls. The subsequent adoption of general deep learning architectures, including convolutional models like DeepLabV3+ and Pyramid Scene Parsing Network, improved pixel-level classification accuracy, yet exposed distinct architectural limitations. Standard convolutional layers operate over square spatial receptive fields, an operational design that fundamentally struggles to capture extended, anisotropic structures such as elongated corridors. Consequently, conventional segmentation models frequently fragment long hallways into disconnected segments while suffering from inadequate geometric-semantic fusion, wherein background noise corrupts room boundaries and causes predicted functional regions to overflow past enclosing walls.
To conquer these structural limitations, the newly formulated EDANet architecture integrates a multi-component processing pipeline centered around a context-enhanced feature extraction backbone. Rather than utilizing standard VGG or residual configurations, the network adopts a DenseNet121 backbone structured into four dense connection blocks alternating with three transition layers. This dense connectivity pattern links multiple layers to enhance feature reuse and facilitate steady gradient propagation, preserving subtle geometric indicators across multi-scale spatial representations. As input images pass through the network, the backbone generates five hierarchical feature stages, extending from shallow representations of 256 by 256 pixels across 64 channels to deep representations measuring 16 by 16 pixels across 1024 channels. To resolve the receptive field limitations of standard square convolutions at this deepest stage, the researchers introduced a dedicated Strip Pooling module. By calculating parallel horizontal and vertical strip averages along feature dimensions, this strip pooling mechanism establishes long-range global context dependencies across perpendicular orientations, directly preventing structural disconnections along lengthy corridors.
Following feature extraction, spatial representations enter a dedicated Spatial Attention Dual-Stream Decoder designed to recover fine geometric details while cleanly separating structural boundaries from functional room identities. Conventional architectural decoders commonly introduce skip connections from early encoder stages to restore spatial resolution, yet these raw connections inevitably channel non-essential background noise, such as furniture patterns, directly into boundary predictions. EDANet mitigates this vulnerability by channeling skip features through a spatial attention mechanism within its four-stage boundary decoding blocks. The spatial attention module compresses incoming feature maps along the channel axis using parallel average pooling and max pooling operations, with max pooling selectively highlighting high-activation structural edges like wall boundaries. After concatenating these two descriptors, the module processes them through a wide seven-by-seven convolutional kernel followed by a Sigmoid activation function to generate a spatial weight map. This attention map dynamically recalibrates encoder features, dampening irrelevant background clutter while sharpening wall corners, window perimeters, and door thresholds.
The complementary half of the decoding system features a parallel semantic decoding stream coupled with an innovative Gated Channel-Spatial Fusion Module, termed GCSFM, which replaces traditional unweighted element-wise feature multiplication. Recognizing that architectural boundaries should impose strict geometric barriers against room semantic propagation, the GCSFM employs boundary predictions to construct an adaptive gating mechanism. The module processes geometric boundary representations through dedicated convolution, batch normalization, and projection operations to yield a spatial gating map with values between zero and one. When applied to incoming room semantic features through a residual formulation, the gate selectively opens along boundary intersections to actively suppress outward semantic leakage while closing across interior room spaces to maintain classification uniformity. To further refine these combined streams, the module concatenates weighted representations from five distinct feature sources and funnels the expanded channel volume into a Squeeze-and-Excitation block. This channel attention unit executes global average pooling and dual fully connected adjustments, dynamically amplifying channel responses associated with unique functional textures while attenuating uninformative background features.
Optimizing this dual-stream architecture requires a balanced loss formulation capable of managing extreme numerical disparities in floor plan pixel distributions, where structural walls occupy far fewer pixels than expansive living areas. The research team implemented a multi-task joint optimization loss function that applies dynamically weighted class balancing based on inverted class occurrence frequencies across the training corpus. Furthermore, total training loss combines boundary segmentation loss and semantic region loss through dynamic weighting parameters, mathematically derived from respective non-background pixel totals to maintain balanced gradient magnitudes across both training objectives. The entire empirical framework was executed on an Ubuntu 22.04 LTS workstation configured with Python 3.10, PyTorch 2.7.1, CUDA 12.8, and cuDNN 9.7.1, utilizing a single NVIDIA GeForce RTX 5080 graphics processing unit with 16 gigabytes of memory. Network training proceeded across 500 epochs with a batch size of 16, utilizing the AdamW optimizer with an initial learning rate of 0.0001, a decoupled weight decay parameter of 0.01, and a Cosine Annealing schedule.
Rigorous empirical benchmarking across diverse architectural layouts validated the system’s operational superiority against leading domain-specific and general computer vision baselines. On the textured Rent3D benchmark, which encompasses 214 original multi-room layouts and 18 circular floor plans split into 179 training and 53 testing images, EDANet achieved a mean Intersection over Union of 78.0 percent, alongside an 87.0 percent Mean Pixel Accuracy and a 95.0 percent Overall Pixel Accuracy. In direct comparison, the domain-tailored FloorNet model recorded an mIoU of 67.0 percent, while the modern Vision Transformer baseline SegFormer reached 74.0 percent. Notably, while SegFormer captured large functional areas effectively, it struggled on slender categories, achieving Class Pixel Accuracy scores of only 67.0 percent for doors and windows and 78.0 percent for hallways. In contrast, EDANet reached 98.0 percent accuracy on doors and windows and 89.0 percent on hallways, demonstrating how strip pooling and spatial gating compensate for boundary adherence failures in self-attention transformer frameworks.
The architecture demonstrated equal robustness when evaluated on the challenging Raster-to-Vector benchmark, an 815-image dataset partitioned into 715 training and 100 testing diagrams that contains exclusively black-and-white structural line drawings devoid of color, furniture iconography, or material textures. Operating without surface indicators, EDANet attained an 81.0 percent mIoU, outperforming both FloorNet at 76.0 percent and SegFormer at 80.0 percent, while achieving a 91.0 percent Mean Pixel Accuracy and a 93.0 percent Overall Pixel Accuracy. In specific structural benchmarks, EDANet maintained commanding advantages over SegFormer, outscoring the transformer 99.0 percent to 77.0 percent on door and window identification and 81.0 percent to 76.0 percent on elongated hallways. Progressive ablation experiments confirmed that while integrating the Gated Channel-Spatial Fusion Module elevated baseline mIoU from 76.0 percent by constraining semantic overflow across open-plan living and kitchen zones, adding Strip Pooling and Spatial Attention provided the decisive topological continuity needed to distinguish master bedrooms from adjacent guest suites.
Detailed confusion matrix evaluations and qualitative visual analyses further illuminated the internal mechanisms driving these performance leaps, illustrating how individual modules cooperatively eliminate common segmentation errors. Under baseline conditions on textured floor plans, small interior functional zones suffered severe misclassification; for instance, closet regions recorded an initial IoU of just 0.18, with substantial pixel proportions mistakenly assigned to bedrooms or living rooms before channel attention recalibration elevated closet identification to 0.56. Similarly, visual inspection revealed that while unaugmented networks produced fragmented corridors and irregular wall breaks, EDANet faithfully reconstructed complete, noise-free spatial layouts aligned with ground truth geometry. Looking toward future industrial deployment, the investigators highlighted two principal development trajectories: formulating automated conversion algorithms to translate raster segmentations into editable vector representations for Building Information Modeling software, and exploring model compression techniques to reduce computational requirements, thereby enabling real-time spatial perception on resource-constrained mobile hardware and autonomous indoor robotic platforms.
Subject of Research: Floor plan semantic segmentation using dual-stream attention networks and explicit boundary constraints
Article Title: Floor Plan Semantic Segmentation Network with Gated Channel-Spatial Fusion and Explicit Boundary Constraint
Article References: Li, Z., Wang, Y., Li, H., Wang, Z., & Yang, Y. (2026). Floor Plan Semantic Segmentation Network with Gated Channel-Spatial Fusion and Explicit Boundary Constraint. Human-Centric Intelligent Systems, 6(2), 187-207. https://doi.org/10.1007/s44230-026-00148-3
Image Credits: AI Generated
DOI: 10.1007/s44230-026-00148-3
Keywords: Floor plan segmentation, Semantic segmentation, EDANet, DenseNet121, Strip pooling, Spatial attention, Dual-stream decoder, Gated channel-spatial fusion, Squeeze-and-Excitation, Boundary constraints, Rent3D dataset, Raster-to-Vector dataset, Mean intersection over union, Architectural drawing analysis, Building Information Modeling, Multi-task learning, Computer vision
Cite Scienmag News
Blake Davidson. (November 18, 2026). Dual-Stream Neural Network Improves Floor Plan Segmentation Through Boundary Constraints. Scienmag. https://scienmag.com/dual-stream-neural-network-improves-floor-plan-segmentation-through-boundary-constraints/
Blake Davidson. "Dual-Stream Neural Network Improves Floor Plan Segmentation Through Boundary Constraints." Scienmag, 18 November 2026, https://scienmag.com/dual-stream-neural-network-improves-floor-plan-segmentation-through-boundary-constraints/. Accessed 20 November 2026.
Blake Davidson. "Dual-Stream Neural Network Improves Floor Plan Segmentation Through Boundary Constraints." Scienmag. November 18, 2026. https://scienmag.com/dual-stream-neural-network-improves-floor-plan-segmentation-through-boundary-constraints/








