Topology-Preserving Data Augmentation for Ring-Type Polygon Annotations

T opology-Pr eserving Data A ugmentation f or Ring-T ype Polygon Annotations Sudip Laudari * 1 Sang Hun Baek 1 Abstract Geometric data augmentation is widely used in segmentation pipelines and typically assumes that polygon annotations represent simply connected regions. Ho we ver , in structured domains such as architectural floorplan analysis, ring-type regions are often encoded as a single cyclic polygon chain connecting outer and inner boundaries. During augmentation, clipping operations may remo ve in- termediate vertices and disrupt this cyclic connec- ti vity , breaking the structural relationship between the boundaries. In this work, we introduce an order-preserving polygon augmentation strategy that performs transformations in mask space and then projects surviving vertices back into index- space to restore adjacency relations. This repair maintains the original trav ersal order of the poly- gon and preserves topological consistency with minimal computational overhead. Experiments demonstrate that the approach reliably restores connectivity , achie ving near-perfect Cyclic Adja- cency Preserv ation (CAP) across both single and compound augmentations. Keyw ords: Data Augmentation, Polygon Segmentation, T opology Preservation, Floorplan Analysis 1. Introduction Segmentation tasks are widely used to extract object bound- aries and spatial structures from images ( Long et al. , 2015 ; Ronneberger et al. , 2015 ). In many datasets, object re- gions are represented using closed polygon boundaries that describe the geometry of the target area. Howe ver , real- world en vironments often contain regions with more com- plex structures, such as corridors, courtyards, walls and spaces with interior voids. These configurations frequently appear in architectural floorplans and other structured vi- sion domains ( Liu et al. , 2018 ; Dodge et al. , 2017 ; Mnih , 1 Independent Researcher . Correspondence to: Sudip Laudari < sudiplaudari@gmail.com > , Sang Hun Baek < shawnbback@gmail.com > . Pr eprint. Mar ch 18, 2026. 2013 ). Accurate polygon representations are therefore es- sential for tasks such as raster -to-vector con version, layout analysis, spatial measurement, and geometric reconstruction ( Lin et al. , 2014 ; Liu et al. , 2018 ; Acuna et al. , 2018 ; Dodge et al. , 2017 ; Girard et al. , 2021 ). T o improve the rob ustness and generalization of segmenta- tion models in such tasks, data augmentation is commonly applied during preprocessing ( Shorten & Khoshgoftaar , 2019 ). Transformations such as rotation, scaling, cropping, and flipping increase data di versity and help reduce o verfit- ting ( Perez & W ang , 2017 ). Modern augmentation frame- works, including Albumentations ( Buslae v et al. , 2020 ) and imgaug ( Jung , 2018 ), support polygon-based annotations and jointly transform images and their vertex coordinates. As a result, these tools ha ve become standard components of many se gmentation pipelines. Howe ver , most augmentation pipelines implicitly assume that each polygon represents a simply connected region. In architectural datasets, regions with interior v oids are often encoded as a single c yclic polygonal chain (see Figure 1 ), where implicit bridge and closure edges connect the outer and inner boundary segments. While af fine transformations preserve c yclic ordering in theory , practical augmentation operations may remov e vertices when polygons are cropped or extend be yond image boundaries. Consequently , the cyclic structure of the polygon can be disrupted, breaking the intended ring representation. This failure mode is often difficult to detect. At the pixel lev el, the augmented region may still appear visually correct, while at the annotation le vel the polygon may no longer pre- serve its intended topology . This can cause a single semantic region to become represented as multiple disconnected frag- ments, introducing inconsistencies in training data, visual- ization, and downstream geometric reasoning ( Mik olajczyk et al. , 2005 ; Benenson et al. , 2019 ; Cordts et al. , 2016 ). In practice, we observe this phenomenon frequently in datasets containing ring-type regions. T o address this issue, we introduce an order-preserving connectivity repair strategy for polygon-based annotation datasets. The approach combines mask-based geometric augmentation with vertex inde x projection and cyclic repair to restore adjacency relations after vertex deletion. The resulting procedure is lightweight, integrates easily into 1 T opology-Preserving P olygon A ugmentation Inner Boundary Bridge Edge Oute r Boundary F igure 1. Failure of ring-type polygon annotations under geometric augmentation. Left: Ground-truth polygon with a narrow bridge connector and inner boundaries are shown in red color and outer boundary in green color . Each corner has a vertex(ke ypoints), and mask (oli ve color) sho ws ring type annotation generated because of wall design. Middle and Right: Representati ve failure cases after augmentation, where the connector is disrupted and the polygon topology is no longer preserved, yielding separate outer mask from outer boundary and inner boundary masks from inner boundary mask. existing augmentation pipelines, and preserves topology- consistent polygon annotations under geometric transforma- tions. 2. Related W ork Data augmentation is a fundamental component of modern segmentation pipelines, impro ving robustness, generaliza- tion, and in variance to geometric transformations ( Shorten & Khoshgoftaar , 2019 ; Perez & W ang , 2017 ). Common operations such as rotation, scaling, translation, cropping, and flipping are widely applied across visual recognition tasks ( Krizhevsk y et al. , 2012 ). In practice, augmentation libraries such as Albumentations ( Buslae v et al. , 2020 ), im- gaug ( Jung , 2018 ), Augmentor ( Bloice et al. , 2017 ), and the recently proposed polygon augmentation library Aug- menT ory ( Ghahremani et al. , 2024 ) provide efficient mecha- nisms for jointly transforming images and their correspond- ing polygon annotations. These frame works are typically used with COCO-style polygon encodings ( Lin et al. , 2014 ), where object regions are represented as ordered v ertex se- quences. Such polygon annotations are widely adopted in large-scale datasets such as COCO ( Lin et al. , 2014 ) and are commonly generated using annotation tools such as La- belMe ( Russell et al. , 2008 ). They are particularly important in structured vision domains including architectural lay- outs, floorplans, maps, and technical drawings ( Dodge et al. , 2017 ; Kale va & K ¨ am ¨ ar ¨ ainen , 2019 ; Zheng et al. , 2020 ), where polygon annotations capture structural relationships between boundaries and spatial regions. While existing augmentation frame works correctly transform v ertex coor- dinates under geometric operations, they primarily focus on spatial consistency and typically assume that polygon connectivity remains v alid after augmentation. In practical machine learning pipelines, augmentation is commonly integrated into training frame works and dataset platforms ( Shorten & Khoshgoftaar , 2019 ; Krizhe vsky et al. , 2012 ; Chen et al. , 2021 ; 2022 ; Zoph et al. , 2020 ). For example, modern detection and segmentation framew orks such as YOLO ( Bochkovskiy et al. , 2020 ; Myshkovsk yi et al. , 2025 ) incorporate online augmentation during train- ing, where images and their annotations are transformed on-the-fly to improve rob ustness. Like wise, recent w ork on polygon-based augmentation such as AugmenT ory ( Ghahre- mani et al. , 2024 ) highlights that polygon and mask annota- tions are routinely augmented using geometric transforma- tions in Y OLO-style instance segmentation models. These systems rely hea vily on geometric transformations applied directly to vertex coordinates or deri ved contours. Ho we ver , when polygons are clipped by image boundaries or crop operations, vertices may be remov ed, potentially disrupting the original adjacency relationships encoded in the polygon sequence. T o mitigate such issues, some approaches perform aug- mentation in mask space rather than directly manipulating polygon vertices. In these methods, polygons are first ras- terized into binary masks, geometric transformations are applied to the masks, and the resulting contours or polygo- nal boundaries are e xtracted afterward ( Shorten & Khosh- goftaar , 2019 ; Liang et al. , 2020 ; Lazarow et al. , 2022 ; Ghahremani et al. , 2024 ). While this strategy simplifies augmentation pipelines, the contour re-extraction process is inherently resolution dependent. As a result, narrow connec- tions such as the bridge between outer and inner boundaries in ring-shaped regions may collapse during rasterization, merging interior holes with the background or fragment- ing the region into multiple components. More generally , theoretical frameworks such as topological data analysis and persistent homology study connectivity and shape in- variants under continuous transformations ( Carlsson , 2009 ; Edelsbrunner et al. , 2008 ), and recent work has introduced topology-aware loss functions to encourage connectivity 2 T opology-Preserving P olygon A ugmentation 𝑃 ! 𝑃 " 𝑃 # 𝑃 #$! 𝑃 #$" 𝑃 % 𝑃 #&! 𝑃 %&" Input Encoding S = O\ I 𝑃 = # (𝑃 ! … … … . . 𝑃 % ) Geometric Augmentatio n Rotate/Crop/ Scale/ T ranslate 𝑀 ! = 𝑇 (𝑀) Index - Space Repair J = # (𝑘 ! … … … . . 𝐾 ' ) 𝑘 ($! # = # (𝑘 ( ( 𝑚𝑜𝑑 #𝑛 )) + 1) Check Successor Cons istency If gap is detec ted à (𝑝 ! ! " , $𝑝 ! (!#$) " ) Projec t i ndic es in Order 𝑃′ ! 𝑃′ # 𝑝 2 𝑃 % Ou t put Polygon 𝑇𝑜𝑝𝑜𝑙𝑜𝑔𝑦 # 𝑃𝑟𝑒𝑠𝑒𝑟𝑣𝑒𝑑 𝑀𝑎𝑠𝑘 $ 𝑇𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛 Outer region - O Inner Boundary Outer Boundary Bridge Edge Bridge Edge Outer Boundary Inner Boundary Inner regio n - I F igure 2. Overvie w of the proposed topology-preserving augmentation pipeline for ring-type polygons. From left to right: the input ring region is encoded as a single cyclic polygon chain P = ( p 1 , . . . , p n ) representing outer and inner boundaries. The polygon is rasterized and geometric augmentation is applied in mask space, M ′ = T ( M ) (e.g., rotate, crop, scale, translate). The surviving polygon indices are then projected in order and successor consistency is checked. If the order is preserved, the polygon remains v alid; otherwise, gaps in the index sequence are repaired using the proposed index-space reconnection method. The resulting polygon ˆ P forms a single closed cyclic chain with preserved ring topology . preservation during model training ( Hu et al. , 2019 ). Ho w- ev er, these methods operate on pixel-le vel predictions and do not address structural corruption that can arise earlier in the pipeline when polygon annotations themselves are altered during augmentation. Despite the widespread use of polygon-based augmentation, the specific problem of cyclic connecti vity failure in single- chain ring encodings has recei ved little attention. Existing augmentation frame works ensure that verte x coordinates are transformed consistently but do not explicitly enforce the adjacency relationships required to maintain ring semantics. In contrast, our work introduces a lightweight repair mech- anism that restores c yclic connectivity after augmentation while remaining compatible with existing polygon-based workflo ws. 3. Method Let P = ( p 1 , p 2 , . . . , p n ) denote a polygon annotation en- coded as a cyclic vertex sequence. In practical datasets, polygon annotations may represent both simply connected regions and ring-type regions containing interior v oids. Ex- amples include architectural structures such as walls, cor- ridors, and exterior spaces surrounding floor boundaries, where the geometry is defined by an ordered sequence of vertices. T o ensure reliable data augmentation preprocessing for com- puter vision pipelines, we introduce an augmentation proce- dure consisting of four stages: (1) rasterizing the polygon into a binary mask, (2) applying geometric augmentation in mask space, (3) projecting surviving vertices back into index space, and (4) restoring cyclic connectivity among the remaining vertices. This procedure preserves the original trav ersal order of the polygon while ensuring that the augmented annotation re- mains topologically consistent. 3.1. Ring-T ype Polygon Encoding W e represent a ring-type region as S = O \ I , where O ⊂ R 2 denotes the outer region boundary and I ⊂ O denotes an interior hole within the re gion. In many annotation formats such as COCO, regions con- taining holes are typically represented using multiple inde- pendent polygons. In contrast, architectural datasets often encode ring-type re gions as a single cyclic polygon chain when floorplan structures are traced or reconstructed. In this representation, both the outer boundary and the inner boundary are stored within a single ordered sequence of vertices. W e therefore represent the polygon as an ordered verte x sequence P = ( p 1 , p 2 , . . . , p n ) , where each verte x p i ∈ R 2 corresponds to a point in the image coordinate system. The sequence is partitioned into two segments: • Outer boundary: B out = ( p 1 , . . . , p L ) • Inner boundary: B in = ( p L +1 , . . . , p n ) These segments are connected through two edges that main- tain the cyclic tra versal of the polygon (see Figure 2 ): 3 T opology-Preserving P olygon A ugmentation e b = ( p L , p L +1 ) where e b denotes the bridge edge connecting the outer and inner boundaries. e c = ( p n , p 1 ) where e c denotes the closure edge completing the cyclic polygon. This encoding allows a single cyclic vertex chain to rep- resent regions containing interior holes while remaining compatible with standard polygon annotation formats. 3.2. Mask-Based Geometric A ugmentation Direct augmentation of polygon vertices may break connec- ti vity when vertices fall outside image boundaries. T o avoid this issue, we perform geometric transformations in mask space. The polygon P is first rasterized into a binary mask M . Geometric augmentation is then applied as M ′ = T ( M ) where T denotes an af fine transformation such as rotation, scaling, translation, or cropping. Each polygon vertex p i is associated with an inde x that preserves the original tra versal order of the polygon. After augmentation, the boundary of the transformed mask M ′ is extracted and used as a geometric reference for verte x projection. For each original verte x index, we search for the nearest boundary point on M ′ that corresponds to the transformed location of the verte x. If the transformed v ertex lies outside the image or outside the valid mask re gion due to cropping or clipping, the origi- nal verte x is discarded. In such cases, the clipping boundary may introduce ne w intersection points that become addi- tional vertices on the augmented polygon. The surviving vertices are then projected back into index space to reconstruct the polygon representation while pre- serving the original cyclic ordering (see Figure 3 ). 3.3. Order -Preserving Connectivity Repair Let J = ( k 1 , k 2 , . . . , k m ) ⊂ { 1 , . . . , n } denote the indices of vertices that remain v alid after augmentation, ordered according to their original cyclic order , where m ≤ n . T o restore cyclic connecti vity , we enforce successor consis- tency among survi ving indices. For each pair of consecutive indices ( k t , k t +1 ) with k m +1 := k 1 , we check whether the original successor relation holds: F igure 3. V ertex projection after mask-based augmentation. The dashed blue polygon shows the original vertex sequence with indices, while the solid orange polygon shows the augmented polygon. Surviving vertices are projected onto the augmented boundary and retain their original indices, preserving the cyclic ordering of the polygon chain. V ertices remo ved by clipping are discarded, allowing the augmented polygon to be reconstructed while maintaining index consistenc y . k t +1 = ( k t (mo d n )) + 1 If this condition is violated, a gap in the original index se- quence indicates that intermediate vertices were remo ved during augmentation. W e then reconnect the surviving ver - tices by enforcing the directed edge ( p ′ k t , p ′ k t +1 ) in the repaired polygon ˆ P . This process preserves the original trav ersal direction while 4 T opology-Preserving P olygon A ugmentation restoring adjacency relations across survi ving vertices. Algorithm 1 Order-Preserving Polygon Augmentation and Connectivity Repair 1: Input: Original polygon P = ( p 1 , p 2 , . . . , p n ) , aug- mentation transform T 2: Output: Repaired cyclic polygon chain ˆ P 3: Rasterize P into binary mask M 4: Apply geometric augmentation: M ′ ← T ( M ) 5: Project original vertices onto M ′ and extract survi ving index set J = ( k 1 , k 2 , . . . , k m ) 6: Sort J according to original cyclic order 7: Initialize ˆ P ← ∅ 8: for t = 1 to m do 9: i ← k t 10: j ← k (( t mo d m )+1) 11: Add directed edge ( p ′ i , p ′ j ) to ˆ P 12: end for 13: return ˆ P 3.4. T opological Consistency Because ring topology is defined by c yclic successor rela- tions in the verte x sequence, maintaining adjacency among surviving indices preserves the structural transitions be- tween B out and B in . The resulting polygon ˆ P therefore forms a single closed cyclic chain that remains consistent with the original ring encoding. 3.5. Implementation and Complexity The repair procedure operates in O ( m ) time, where m is the number of surviving vertices after augmentation ( m ≤ n ). All operations are performed in index space and there- fore introduce negligible computational ov erhead compared with standard geometric augmentation operations. Because the method av oids contour reconstruction or polygon re- detection, it remains lightweight and can be easily integrated into existing augmentation pipelines. 4. Experimental Setup Our ev aluation focuses on the structural correctness of poly- gon annotations under geometric data augmentation. In particular , we analyze whether augmentation pipelines pre- serve the cyclic topology of ring-type polygons and whether the proposed repair strategy restores structural consistenc y when verte x deletion occurs. Unlike con ventional segmentation studies, our e xperiments ev aluate annotation-lev el topology rather than downstream segmentation accurac y . This allo ws us to directly measure how augmentation procedures af fect the structural integrity of polygon representations. All augmentation methods are applied using identical pa- rameter ranges. Metrics are computed per semantic class and av eraged across samples. 4.1. A ugmentation Setup A ugmentation Parameters: W e e valuate commonly used geometric augmentations in segmentation pipelines, includ- ing rotation, scaling, translation, cropping, and horizontal flipping. The parameter ranges follow standard augmenta- tion practices and are summarized in T able 1 . Color-based augmentations are excluded since they do not alter polygon geometry . T able 1. Geometric augmentation parameters evaluated in this study . A ugmentation Parameter Range Rotation Angle [ − 30 ◦ , 30 ◦ ] Scaling (in/out) Scale factor [0 . 7 , 1 . 3] Cropping Crop scale [0 . 6 , 1 . 0] Rotation + Cropping Angle / scale [ − 30 ◦ , 30 ◦ ] , [0 . 6 , 1 . 0] T ranslation Shift ratio [ − 0 . 1 , 0 . 1] (x, y) Flipping H / V flip Probability = 0 . 5 Baselines: W e compare our proposed method against widely used augmentation workflo ws to assess cyclic topol- ogy preservation under practical conditions. • Y OLOv11 Pipeline ( Ultralytics , 2024 ). Represents a typical training-time augmentation strate gy used in modern detection and segmentation frame works. Geo- metric transformations are applied during model train- ing, and polygon annotations are transformed jointly with images. Howe ver , no explicit validation of c yclic verte x adjacency is performed after v ertex remov al. • Roboflow A ugmentation ( Roboflow , Inc. , 2023 ). Represents a preprocessing-based augmentation work- flow widely used in dataset preparation platforms. Polygon vertices are transformed according to geo- metric mappings, but cyclic index adjacency is not enforced when vertices are remov ed due to clipping or cropping. • Ours (Proposed). Uses the Albumentations geometric engine ( Buslaev et al. , 2020 ) for spatial transforma- tions and applies our inde x-preserving repair module to restore cyclic adjacenc y among survi ving vertices. 4.2. Evaluation Metric: Cyclic Adjacency Preserv ation (CAP) T o quantify structural integrity of polygon annotations, we introduce the Cyclic Adjacency Pr eservation (CAP) metric. 5 T opology-Preserving P olygon A ugmentation Let the original polygon be defined by the cyclic inde x set I = { 1 , . . . , n } with successor function succ( i ) = ( i mod n ) + 1 . After augmentation and boundary clipping, let J = ( k 1 , k 2 , . . . , k m ) denote the ordered set of surviving v ertex indices, where m ≤ n and k m +1 := k 1 . CAP measures the proportion of survi ving adjacency rela- tions that remain consistent with the original cyclic order: CAP( P → P ′ ) = 1 m m X t =1 I [ k t +1 = succ( k t )] , (1) where I [ · ] denotes the indicator function. The metric CAP ∈ [0 , 1] equals 1 . 0 only when all surviving vertices maintain their original cyclic successor relations. Unlike geometric ov erlap metrics such as mIoU, CAP ev al- uates preservation of the polygon’ s connecti vity structure rather than spatial ov erlap. Importantly , CAP does not penalize legitimate verte x re- mov al caused by cropping. Instead, it detects discontinuities in the remaining v ertex sequence that break cyclic connec- tivity . For a dataset containing N ring-type polygon instances, we report the mean Cyclic Adjacency Preserv ation: CAP = 1 N N X i =1 CAP( P i → P ′ i ) . (2) 5. Results The quantitativ e ev aluation of cyclic topology preservation under rotation augmentation is summarized in T able 2 . Stan- dard augmentation pipelines frequently introduce cyclic ad- jacency violations because v ertices may be remov ed during clipping or cropping operations. When intermediate vertices disappear , the successor relationships encoded in the origi- nal polygon sequence are disrupted, producing fragmented polygon chains and inconsistent topology . The proposed repair mechanism addresses this issue by restoring adjacency relations among survi ving v ertices through successor consistency in inde x space. As reflected in T able 2 , the proposed method achieves CAP v alues close to 1 . 0 , indicating near -perfect preserv ation of the original cyclic polygon structure after augmentation. In contrast, baseline pipelines e xhibit substantially lo wer CAP values, suggesting that con ventional augmentation strategies do not reliably maintain the structural inte grity of ring-type poly- gons. These failures occur because geometric operations such as cropping or boundary clipping remo ve intermedi- ate vertices without updating the cyclic adjacenc y relations encoded in the polygon sequence. In addition, some aug- mentation pipelines reconstruct polygons from rasterized masks, which can alter v ertex ordering or introduce dense contour points that further disrupt the original c yclic struc- ture. T able 2. Mean Cyclic Adjacency Preserv ation (CAP) under rota- tion augmentation. Higher values indicate better preserv ation of cyclic polygon topology . Method CAP ( ↑ ) Y OLO Augmentation 0.3278 Roboflow 0.5497 Ours 0.9758 T able 3. Mean Cyclic Adjacenc y Preservation (CAP) across geo- metric augmentations using the proposed repair strategy . A ugmentation CAP ( ↑ ) Rotation 0.9774 Cropping 0.9840 Scaling 0.9827 Flip 0.9869 Rotation + Cropping 0.9748 Qualitativ e comparisons across augmentation pipelines further illustrate these differences (Figures 4 – 6 ). Under Roboflow augmentation, clipping during geometric transfor- mations removes the bridge connector between the outer and inner boundaries (Figure 4 ). As a result, the polygon chain becomes fragmented and the annotation splits into separate masks representing the outer and inner regions, breaking the intended ring topology . A different beha vior is observed in the YOLO training aug- mentation pipeline (Figure 5 ). Although the topology col- lapse seen in Roboflow does not occur , the augmentation introduces an excessi ve number of boundary vertices along the outer contour . This produces an inflated polygon rep- resentation with more keypoints than required, altering the original boundary structure. Such behavior likely arises from raster-to-v ector contour extraction commonly used in augmentation framew orks. Instead of directly transforming the original vertex indices, the augmented mask is con verted back into a polygon by re-extracting contours, resulting in a dense sequence of pixel-aligned v ertices that deviates from the sparse and structurally consistent representation of the ground truth. In contrast, the proposed repair strategy preserv es the cyclic ordering of vertices and reconstructs the polygon chain af- ter augmentation by restoring missing successor relations 6 T opology-Preserving P olygon A ugmentation F igure 4. Qualitativ e results using Roboflow augmentation. The left panel shows the ground truth polygon with a ring structure. The middle and right panels illustrate failure cases after augmentation, where clipping and rotation remove the bridge connection between the outer and inner boundaries. As a result, the original ring polygon becomes fragmented and is incorrectly represented as two independent polygons corresponding to the outer and inner regions. F igure 5. Examples illustrating topology failure under Y OLO training augmentation with default parameters. The upper row sho ws the ground truth ring-type polygon annotations, while the lo wer ro w shows the augmented outputs. After augmentation, the interior hole is lost and the region is incorrectly represented as a single lar ge polygon, breaking the intended ring topology . 7 T opology-Preserving P olygon A ugmentation F igure 6. Results produced by the proposed topology-preserving augmentation. The upper ro w shows the ground truth ring-type polygon annotations, while the lower ro w presents the augmented outputs generated by our method. The red circles highlight locations where the bridge connections between the outer and inner boundaries are preserv ed after augmentation, maintaining the cyclic polygon chain and prev enting topology fragmentation. (Figure 6 ). The resulting polygons therefore maintain a single closed cyclic representation while preserving the cor- rect outer–inner boundary relationships without introducing redundant vertices. Performance across multiple geometric transformations is summarized in T able 3 . Cyclic adjacenc y preservation re- mains close to unity for all augmentation types, including compound transformations such as rotation combined with cropping. These results indicate that the proposed repair strategy remains robust under a wide range of geometric transformations commonly used in polygon-based segmen- tation pipelines. Overall, the results demonstrate that enforcing successor consistency in index space effecti vely preserves polygon topology during geometric augmentation and enables reli- able inte gration of augmentation pipelines in polygon-based segmentation workflo ws. Effect of T opology Preser vation on Segmentation T rain- ing: T o assess the practical impact of topology preser - vation, we train segmentation models using datasets aug- mented with dif ferent pipelines and compare standard aug- mentation with the proposed topology-preserving strategy . T wo segmentation architectures are e valuated: Y OLOv11- Seg and Mask R-CNN. All models are trained with identical hyperparameters and e valuated using mean Intersection o ver Union (mIoU). Experiments are conducted on a dataset of approximately 600 K orean floorplan images with polygon-based segmen- tation annotations for architectural regions. The dataset is divided into training, v alidation, and test sets using a 70%, 20%, and 10% split. Geometric augmentation is applied only to the training set to increase data div ersity . For each training image, fi ve aug- mented samples are generated using random transformations sampled from the parameter ranges listed in T able 1 . The same augmentation settings are used for both baseline and topology-preserving pipelines to ensure a fair comparison. T able 4. Se gmentation performance using dif ferent augmentation strategies. Method mIoU ↑ Y OLOv11-Seg (Std. Aug.) 88.4 Y OLOv11-Seg (Ours) 90.8 Mask R-CNN (Std. Aug.) 85.2 Mask R-CNN (Ours) 87.0 The dataset used in this study is publicly av ailable through the AI Hub platform (https://aihub .or .kr/). 8 T opology-Preserving P olygon A ugmentation The results indicate that preserving polygon topology during augmentation improv es segmentation performance. Models trained with topology-consistent annotations achiev e higher mIoU and produce more accurate boundary predictions, suggesting that eliminating structural inconsistencies in aug- mented annotations allows the models to learn geometric boundaries more reliably . 6. Code A vailability The implementation is publicly av ailable as a Python pack- age and can be installed via: pip install polyaug The source code and experimental scripts are a vailable at: https://github.com/Laudarisd/polyaug 7. Conclusion This w ork studies the structural impact of geometric data augmentation on polygon-based segmentation annotations. Although geometric transformations theoretically preserve polygon geometry , practical augmentation pipelines often remov e vertices during clipping or cropping, breaking the cyclic adjacenc y relations that define ring-type regions. Using the proposed Cyclic Adjacenc y Preservation (CAP) metric, our experiments sho w that common augmentation workflo ws frequently introduce structural discontinuities in polygon annotations. In contrast, enforcing index-consistent connectivity preserves the cyclic structure of polygons across a wide range of geometric transformations. These findings highlight the importance of maintaining annotation-lev el topology during data augmentation. While motiv ated by ring-type re gions commonly found in archi- tectural datasets, the proposed index-aw are repair principle can be applied to general polygon annotations where vertex ordering encodes structural relationships. Preserving such structural consistency can impro ve the reliability of polygon- based segmentation pipelines and downstream geometric learning tasks. References Acuna, D., Ling, H., Kar , A., and Fidler , S. Efficient inter - activ e annotation of segmentation datasets with polygon- rnn++. In Pr oceedings of the IEEE conference on Com- puter V ision and P attern Recognition , pp. 859–868, 2018. Benenson, R., Popo v , S., and Ferrari, V . Large-scale inter - activ e object segmentation with human annotators. In Pr oceedings of the IEEE/CVF conference on computer vision and pattern r ecognition , pp. 11700–11709, 2019. Bloice, M. D., Stock er , C., and Holzinger , A. Augmen- tor: an image augmentation library for machine learning. arXiv pr eprint arXiv:1708.04680 , 2017. Bochkovskiy , A., W ang, C.-Y ., and Liao, H.-Y . M. Y olov4: Optimal speed and accurac y of object detection. arXiv pr eprint arXiv:2004.10934 , 2020. Buslaev , A., Iglovik ov , V . I., Khvedchenya, E., P arinov , A., Druzhinin, M., and Kalinin, A. A. Albumentations: Fast and flexible image augmentations. Information , 11(2): 125, 2020. doi: 10.3390/info11020125. Carlsson, G. T opology and data. Bulletin of the American Mathematical Society , 46(2):255–308, 2009. doi: 10. 1090/S0273- 0979- 09- 01249- X. Chen, Y ., Li, Y ., Kong, T ., Qi, L., Chu, R., Li, L., and Jia, J. Scale-aware automatic augmentation for object detection. In Pr oceedings of the IEEE/CVF confer ence on computer vision and pattern r ecognition , pp. 9563–9572, 2021. Chen, Y ., Zhang, P ., Kong, T ., Li, Y ., Zhang, X., Qi, L., Sun, J., and Jia, J. Scale-aware automatic augmentations for object detection with dynamic training. IEEE T ransac- tions on P attern Analysis and Machine Intelligence , 45 (2):2367–2383, 2022. Cordts, M., Omran, M., Ramos, S., Rehfeld, T ., Enzweiler , M., Benenson, R., Franke, U., Roth, S., and Schiele, B. The cityscapes dataset for semantic urban scene un- derstanding. In Pr oceedings of the IEEE conference on computer vision and pattern r ecognition , pp. 3213–3223, 2016. Dodge, S., Xu, J., and Stenger, B. P arsing floor plan im- ages. In 2017 Fifteenth IAPR international confer ence on machine vision applications (MV A) , pp. 358–361. IEEE, 2017. Edelsbrunner , H., Harer , J., et al. Persistent homology-a surve y . Contemporary mathematics , 453(26):257–282, 2008. Ghahremani, T ., Hoseyni, M., Ahmadi, M. J., Mehrabi, P ., and Nikoof ard, A. Augmentory: A fast and flex- ible polygon augmentation library . arXiv preprint arXiv:2405.04442 , 2024. Girard, N., Smirnov , D., Solomon, J., and T arabalka, Y . Polygonal building e xtraction by frame field learning. In Pr oceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition , pp. 5891–5900, 2021. Hu, X., Li, F ., Samaras, D., and Chen, C. T opology- preserving deep image segmentation. Advances in neural information pr ocessing systems , 32, 2019. 9 T opology-Preserving P olygon A ugmentation Jung, A. imgaug, 2018. https://github .com/aleju/imgaug. Kalev a, A. and K ¨ am ¨ ar ¨ ainen, J.-K. Cubicasa5k: A dataset and an improv ed multi-task model for floorplan image analysis. In Scandinavian Conference on Image Analysis , 2019. Krizhevsk y , A., Sutske ver , I., and Hinton, G. E. Imagenet classification with deep con volutional neural networks. Advances in neural information processing systems , 25, 2012. Lazarow , J., Xu, W ., and T u, Z. Instance segmentation with mask-supervised polygonal regression transformers. In IEEE Computer Society Confer ence on Computer V ision and P attern Recognition , 2022. Liang, J., Homayounfar , N., Ma, W .-C., Xiong, Y ., Hu, R., and Urtasun, R. Polytransform: Deep polygon trans- former for instance segmentation. In Pr oceedings of the IEEE/CVF confer ence on computer vision and pattern r ecognition , pp. 9131–9140, 2020. Lin, T .-Y ., Maire, M., Belongie, S., Hays, J., Perona, P ., Ra- manan, D., Doll ´ ar , P ., and Zitnick, C. L. Microsoft coco: Common objects in context. In European Confer ence on Computer V ision (ECCV) , pp. 740–755, 2014. doi: 10.1007/978- 3- 319- 10602- 1 48. Liu, C., W u, J., and Furukawa, Y . Floornet: A unified framew ork for floorplan reconstruction from 3d scans. In Pr oceedings of the Eur opean confer ence on computer vision (ECCV) , pp. 201–217, 2018. Long, J., Shelhamer , E., and Darrell, T . Fully con volutional networks for semantic se gmentation. In Pr oceedings of the IEEE conference on computer vision and pattern r ecognition , pp. 3431–3440, 2015. Mikolajczyk, K., T uytelaars, T ., Schmid, C., Zisserman, A., Matas, J., Schaf falitzk y , F ., Kadir, T ., and Gool, L. V . A comparison of affine re gion detectors: Mikolajczyk et al. International journal of computer vision , 65(1):43–72, 2005. Mnih, V . Machine learning for aerial image labeling . Uni- versity of T oronto (Canada), 2013. Myshkovsk yi, Y ., Nazarke vych, M., Vysotska, V ., and Y urynets, R. Exponential data augmentation methods for improving yolo performance in computer vision tasks. 2025. Perez, L. and W ang, J. The ef fectiveness of data augmen- tation in image classification using deep learning. arXiv pr eprint arXiv:1712.04621 , 2017. Roboflow , Inc. Roboflow: Annotation and augmentation platform. https://roboflow.com , 2023. Ronneberger , O., Fischer , P ., and Brox, T . U-net: Con volu- tional networks for biomedical image se gmentation. In In- ternational Confer ence on Medical image computing and computer-assisted intervention , pp. 234–241. Springer , 2015. Russell, B. C., T orralba, A., Murphy , K. P ., and Freeman, W . T . Labelme: a database and web-based tool for image annotation. International journal of computer vision , 77 (1):157–173, 2008. Shorten, C. and Khoshgoftaar, T . M. A survey on image data augmentation for deep learning. J ournal of Big Data , 6(1):60, 2019. doi: 10.1186/s40537- 019- 0197- 0. Ultralytics. Y olo v11: Real-time object detection. https: //github.com/ultralytics/ultralytics , 2024. Zheng, J., Li, Y ., et al. Structured3d: A large photo-realistic dataset for structured 3d modeling. In ECCV , 2020. Zoph, B., Cub uk, E. D., Ghiasi, G., Lin, T .-Y ., Shlens, J., and Le, Q. V . Learning data augmentation strategies for object detection. In European confer ence on computer vision , pp. 566–583. Springer , 2020. 10