Roboticists investigating autonomous navigation have demonstrated that combining deep learning proposal networks with graph-based geometric solvers significantly improves rectangular landmark detection across robotic environments. In structured environments, rectangular primitives such as doors, windows, displays, and signs provide planar geometry and four-corner topology that remain robust across viewpoints. While conventional localization systems frequently rely on point features that struggle with repetitive textures or perceptual drift, high-level geometric elements supply dependable structural anchors. A research team led by Marco David Arias, Zhenhao Zhang, Jiaxi Huo, and Paloma de la Puente investigated this capability by evaluating extensions of the Rectangle Intersection-based Detector using Graphs and Elongation, known as RIDGE. Their experiments demonstrate that guiding geometric extraction via learned proposals curtails false detections while preserving metric vertex accuracy for mobile robot positioning.
Within Simultaneous Localization and Mapping and visual robot localization, accurate rectangle extraction establishes stable geometric references that mitigate spatial ambiguity. Standard object detectors typically predict loose bounding boxes capturing background clutter alongside targets, limiting direct metric utility in robotics without further refinement. In contrast, physical rectangles maintain parallel opposite edges and orthogonal corners, even though perspective projection transforms their observed image contours into arbitrary quadrilaterals. Previous research demonstrated that integrating rectangular landmarks into Adaptive Monte Carlo Localization routines significantly curtails longitudinal translation uncertainty in uniform corridors. Furthermore, in symmetric layouts featuring identical rooms where range-only sensors produce catastrophic localization errors, recognizing individual rectangular entities supplies the visual distinctiveness needed to disambiguate the robot’s pose. Exploiting these high-level structures reliably requires specialized computational frameworks balancing geometric precision with processing speed.
The baseline RIDGE method was originally conceived as a graph-based algorithm for extracting quadrilaterals from images without learned semantic priors. In this formulation, a probabilistic Fast Line Detector extracts linear segments that are extended to discover candidate intersection points. Recording these coordinates within an adjacency matrix yields four-node cycles in the connectivity graph corresponding to quadrilateral hypotheses. Although RIDGE operates effectively in uncluttered scenes containing isolated rectangular objects, evaluations revealed significant vulnerabilities in complex environments. The underlying line frontend relies on rigid parameters, including a minimum length threshold of fifteen pixels, which fails across varying scales and resolutions. Furthermore, intricate textures, shadows, and low-contrast boundaries produce spurious line fragments, generating excessive intersections and degraded corner localization that compromise landmark recognition.
To overcome RIDGE’s sensitivity to fragmented edges and ambient noise, the researchers designed RIDGE-AirLine, replacing the classical Fast Line Detector with the hybrid AirLine frontend. This architecture employs deep convolutional backbones, including DexiNed and UNet models, to extract structural representations resilient to noise, illumination shifts, and textured backgrounds. Following feature extraction, AirLine applies Contextual Region Growing, dynamically tuning detection sensitivity based on local visual context and reconnecting broken segments along orientation-aligned growth paths. Additionally, the hybrid detector integrates a multi-angle orientation mechanism using directional convolutional filters to highlight linear structures aligned with dominant environmental axes, such as architectural verticals and horizontals. Discarding irrelevant oblique textures and reinforcing salient structural boundaries delivers cleaner linear inputs to subsequent elongation and intersection routines, yielding greater candidate stability across complex scenes.
Targeting computational overhead and spurious intersections, the authors developed RIDGE-ROI to restrict geometric search to promising visual subregions. Rather than exhaustively computing line extensions across the entire image grid, RIDGE-ROI initiates processing through contour extraction, adaptive luminance thresholding, and geometric filtering that isolates regions displaying four convex sides. The baseline RIDGE algorithm then operates exclusively within each extracted region of interest, substantially narrowing the search area and omitting background clutter. Furthermore, whereas original RIDGE applied Non-Maximum Suppression strictly to adjacent intersection points, RIDGE-ROI introduces Non-Maximum Suppression at the rectangle level by evaluating bounding box Intersection-over-Union. Whenever two candidate rectangles exceed an overlap threshold, the system automatically retains solely the largest detected entity, successfully eliminating nested boundary duplicates and producing a concise set of distinct planar landmarks.
Recognizing the benefits of deep learning for semantic categorization, the investigators formulated RIDGE-YOLO, an integrated framework decoupling semantic localization from metric corner extraction. A lightweight YOLOv8n convolutional model processes incoming frames to rapidly locate candidate rectangular elements, outputting bounding box coordinates and semantic class labels. The geometric RIDGE pipeline subsequently analyzes internal geometry exclusively within these proposal crops, focusing computation on high-probability regions. Limiting line detection and graph-based polygon reconstruction to local windows dramatically reduces processing time and filters out background clutter before candidate generation begins. Moreover, because the initial detector assigns explicit semantic categories—distinguishing wall displays from doors or cabinetry—the architecture provides categorical metadata that robotic data association algorithms can exploit, preventing incorrect landmark associations when the mobile platform navigates structurally similar indoor rooms.
Beyond pipeline modifications, the researchers introduced YOLO-BR, an exploratory boundary refinement technique replacing intersection-based corner graphs with a directed compatibility graph of supporting lines. Operating within padded crops defined by YOLO bounding boxes, the Fast Line Detector extracts finite segments, utilizing only original detected lengths without artificial elongation. Each segment defines a supporting line in homogeneous coordinates, oriented so its normal vector points toward the proposal box center. The algorithm constructs a directed graph where supporting lines serve as nodes, connecting pairs that intersect within the proposal crop and satisfy a positive signed determinant condition to ensure consistent cyclic traversal. Excluding near-parallel line pairs, the system extracts directed four-line cycles, determining quadrilateral candidate vertices via homogeneous cross products without requiring lines to conform to predefined horizontal or vertical categories.
To score quadrilateral hypotheses generated by four-line cycles, YOLO-BR evaluates observed boundary support and spatial alignment with the detector proposal. For each candidate side, detected line segments within a perpendicular distance band of 1.414 crop pixels are projected onto the boundary to establish overlapping intervals. The formulation computes direction-weighted coverage by integrating directional agreement along the edge length, requiring coverage of at least 0.5. To penalize oversegmentation where detected lines extend beyond candidate corners, an interval evidence metric divides supported length by total observed span. A candidate’s final joint score balances mean interval evidence across all four edges with its proposal bounding box Intersection-over-Union affinity. Crucially, YOLO-BR selects at most one supported quadrilateral per proposal and rejects any proposal lacking an admissible candidate, avoiding unsupported boundaries.
To address the lack of pixel-accurate vertex benchmarks for geometric rectangle evaluation, the authors created and publicly released two novel annotated datasets. The synthetic IsaacSim50 benchmark features fifty multi-view images captured at ten frames per second across simulated hospital and office environments in IsaacSim. Employing Omniverse Replicator, the authors incorporated randomized objects, lighting variations, and occlusions, providing seventy-six evaluated rectangle instances. To train their detector, they generated 750 randomized scenes, training a YOLOv8n-seg model over 100 epochs at 640 by 640 resolution, achieving confusion matrix scores of 0.88 for paintings, 0.85 for boxes, 0.83 for cardboard, and 0.81 for outlets. For real-world benchmarking, the team constructed RobotHome80, selecting eighty RGB-D images from Robot@Home containing 146 rectangular instances, annotated semi-automatically with Roboflow Annotate and SAM 2 before manual vertex verification.
Comprehensive experiments conducted across RobotHome80, IsaacSim50, and the document-centric SmartDoc90 benchmark demonstrated distinctive operational tradeoffs among the methods, which were also evaluated against the external M-LSD learned line baseline. On RobotHome80, RIDGE-YOLO achieved the highest precision and F1 score, whereas YOLO-BR attained the highest recall. Conversely, on synthetic IsaacSim50 images, YOLO-BR secured top values across precision, recall, and F1 score simultaneously, highlighting the utility of strict boundary evidence in clean environments. Analyzing performance across varying Intersection-over-Union thresholds revealed that RIDGE-YOLO maintained the most stable precision and recall under demanding overlap criteria. RIDGE-AirLine performed competitively at lower thresholds but experienced sharp metric declines at higher requirements. Average Precision values reflected real-world visual challenges, as RIDGE-YOLO reached 0.800 on synthetic images but fell to 0.741 on RobotHome80 due to complex illumination and textures.
Computational efficiency evaluations conducted on an Intel Core i7-10700 system equipped with 32 gigabytes of RAM and an NVIDIA GeForce RTX 2070 SUPER GPU revealed that execution times depend heavily on scene complexity and image resolution. Processing durations were longer on synthetic IsaacSim50 frames than on real RobotHome80 images, resulting in lower frame rates. Higher rendering resolution in IsaacSim50 prompted line detectors to extract substantially more edges and textures, multiplying candidate vertices and graph evaluation overhead. In contrast, RobotHome80 frames yielded fewer detected lines, lowering computational load and raising achieved frame rates. Qualitative assessments on autonomous driving images from KITTI and complex domestic interiors confirmed that while RIDGE-YOLO effectively localizes signs and furniture, occasional failure modes arise, such as incorporating auxiliary text into signs or generating false positives from floor textures and furniture legs.
The experimental findings confirm that coupling deep learning proposals with geometric graph methods establishes a robust framework for visual landmark detection in robotics. Restricting geometric analysis to semantically localized regions prevents the combinatorial explosion of false intersections that previously hampered classical detectors in cluttered settings. Integrating semantic labels with precise vertex coordinates provides dual-stage verification that streamlines data association during particle filter localization and visual SLAM. Nonetheless, notable limitations persist in unconstrained operational environments. Boundary extraction degrades under severe physical occlusions, extremely thin borders, specular reflections, and extreme perspective distortions where quadrilaterals collapse toward degenerate forms. Furthermore, because existing implementations process frames independently without enforcing temporal consistency across video sequences, future investigations will prioritize multi-frame tracking and temporal filtering to maintain geometric consistency and long-term stability during continuous mobile robot navigation.
Subject of Research: Graph-based geometric methods and deep learning frameworks for rectangle landmark detection in robotics
Article Title: Towards novel reliable landmarks in robotics: evaluation of graph-based methods for rectangle detection in images
Article References: Arias, M. D., Zhang, Z., Huo, J., & de la Puente, P. (2026). Towards novel reliable landmarks in robotics: evaluation of graph-based methods for rectangle detection in images. International Journal of Intelligent Robotics and Applications. https://doi.org/10.1007/s41315-026-00610-2
Image Credits: AI Generated
DOI: 10.1007/s41315-026-00610-2
Keywords: robotics localization, rectangle detection, RIDGE algorithm, YOLO-BR, RIDGE-YOLO, RIDGE-AirLine, RIDGE-ROI, simultaneous localization and mapping, IsaacSim50 dataset, RobotHome80 dataset, SmartDoc90, supporting-line graph, M-LSD baseline, geometric landmark extraction, fast line detector, contextual region growing, boundary refinement, planar geometry
News Source: Denise Maddox. (November 18, 2026). Hybrid Graph-Based Detectors Boost Rectangle Landmark Accuracy for Mobile Robot Localization. Scienmag.



