HoliTracer: Direct Vectorization in RSI

• HoliTracer is a framework that directly vectorizes entire large-scale remote sensing images, preserving global context and reducing patch boundary fragmentation.
• Its Context Attention Net (CAN) uses multi-scale image pyramids and local-to-global attention to enhance segmentation accuracy for buildings, water bodies, and roads.
• The two-stage vectorization pipeline, combining Mask Contour Reformer (MCR) and Polygon Sequence Tracer (PST), improves polygon quality by refining contours and vertex positions.

HoliTracer is a framework for direct vectorization of large-size remote sensing imagery (RSI), introduced to holistically extract vectorized geographic objects from large-size RSI without patch stitching. It targets buildings, water bodies, and roads in very-high-resolution imagery whose spatial extent commonly exceeds $10{,}000\times 10{,}000$ pixels, where patch-based processing can discard global scene context and produce fragmented vector outputs at patch boundaries. The framework combines a context-aware segmentation module, termed Context Attention Net (CAN), with a two-stage vectorization pipeline composed of the Mask Contour Reformer (MCR) and the Polygon Sequence Tracer (PST) (Wang et al., 22 Jul 2025).

1. Problem setting and design rationale

HoliTracer is motivated by the observation that modern very-high-resolution RSIs easily exceed $10{,}000\times 10{,}000$ pixels, whereas most existing vectorization methods are constrained to small patches such as $512\times 512$. In the motivating formulation, patch-based cropping discards global scene context, which can lead to misclassification of objects that require large-scale cues, such as confusing building rooftops with parking lots, and also yields fragmented vector outputs at patch boundaries. The problem is compounded by the fact that buildings, water bodies, and roads vary drastically in scale and shape, complicating a unified representation across categories (Wang et al., 22 Jul 2025).

The framework is organized around four stated contributions. First, it provides a unified framework to directly vectorize entire large-size RSIs without patch stitching. Second, CAN employs a local-to-global attention mechanism on a multi-scale image pyramid to recover lost context and improve segmentation completeness. Third, vectorization is handled by a robust two-stage pipeline in which MCR simplifies and uniformly reconstructs polygonal contours from segmentation masks, while PST refines point positions and predicts true vertices through iterative offset regression and transformer-based classification with geometric angle features. Fourth, the method is reported to outperform state-of-the-art patch-based and large-scale segmentation baselines on building, water body, and road datasets (Wang et al., 22 Jul 2025).

A plausible implication is that HoliTracer is best understood not as a pure contour extraction method, but as an end-to-end large-scene mapping pipeline in which segmentation quality and polygon tracing are co-dependent stages.

2. Context Attention Net for large-scene segmentation

CAN addresses the contextual deficiencies of patch-limited processing through a multi-scale image pyramid. Given an input RSI $I$ of size $>10{,}000\times 10{,}000$, the method constructs downsampled versions with rates $d_1=1<d_2<\dots<d_n$, with examples $\{1,3,6\}$ or $\{1,5,10\}$. A fixed-size window, such as $512\times 512$, is slid across each scale to yield aligned patch sets $\{I_1,I_2,\dots,I_n\}$, where $10{,}000\times 10{,}000$0 captures fine detail and higher-scale inputs provide increasingly global context (Wang et al., 22 Jul 2025).

Feature extraction is performed with a Swin-L transformer pretrained with Skysense weights. Each scale is encoded as

$10{,}000\times 10{,}000$1

and the upper-scale encoders share weights with the bottom scale. Cross-scale fusion is then performed by local-to-global attention: $10{,}000\times 10{,}000$2

$10{,}000\times 10{,}000$3

The fused representation is

$10{,}000\times 10{,}000$4

This fused tensor is decoded by a UPerNet segmentation head into per-pixel class logits, optimized with cross-entropy loss

$10{,}000\times 10{,}000$5

where $10{,}000\times 10{,}000$6 is the number of pixels and $10{,}000\times 10{,}000$7 the number of classes (Wang et al., 22 Jul 2025).

The architectural significance of CAN lies in the explicit separation between local detail and large-scale semantic support. The reported qualitative analysis states that holistic segmentation with CAN recovers rooftop and building context and avoids misclassification of parking areas, which suggests that the model’s central contribution is not merely scale augmentation, but aligned cross-scale contextual conditioning over the entire large tile (Wang et al., 22 Jul 2025).

3. Mask Contour Reformer and polygon reconstruction

MCR converts segmentation masks into polygon sequences suitable for supervised vertex tracing. Its pipeline begins by extracting a raw contour $10{,}000\times 10{,}000$8 from the binary segmentation mask using TC89-KCOS. The contour is then simplified with Douglas–Peucker, using tolerance $10{,}000\times 10{,}000$9, to obtain $512\times 512$0. From this simplified contour, a uniformly spaced point sequence $512\times 512$1 is reconstructed by interpolating along each edge $512\times 512$2 every $512\times 512$3 pixels, with

$512\times 512$4

and $512\times 512$5 inserted points on that segment (Wang et al., 22 Jul 2025).

During training, this reconstructed sequence is aligned with the ground-truth polygon $512\times 512$6. For each $512\times 512$7, the nearest reconstructed point is selected,

$512\times 512$8

matched points are marked as vertices with $512\times 512$9, and the remaining points receive $I$0. Between consecutive matched vertices, linear interpolation produces $I$1 with $I$2, yielding training pairs $I$3. The key correspondence constraint is

$I$4

which guarantees a one-to-one correspondence for supervised vertex learning (Wang et al., 22 Jul 2025).

This stage performs two distinct functions. First, it suppresses the irregularity of raw mask contours by simplification. Second, it imposes a uniform sampling structure that makes subsequent sequence learning well-posed. The reported qualitative analysis states that MCR yields clean, uniformly sampled contours and prevents boundary fragmentation across large tiles, which indicates that its role is as much representational as geometric (Wang et al., 22 Jul 2025).

4. Polygon Sequence Tracer and vertex inference

PST takes the reconstructed polygon $I$5 and performs point refinement followed by vertex classification. For each point $I$6, local image features $I$7 are extracted from the frozen encoder feature map, and tokens of the form $I$8 are passed into a transformer-based offset regressor that predicts offsets $I$9. Refinement is iterative: $>10{,}000\times 10{,}000$0 After the final refinement step, the refined sequence $>10{,}000\times 10{,}000$1 is used to compute corner angles for $>10{,}000\times 10{,}000$2: $>10{,}000\times 10{,}000$3 where $>10{,}000\times 10{,}000$4 with cyclic indexing. The angle tuple $>10{,}000\times 10{,}000$5 is inserted as a geometric token, and the concatenated descriptor is fed to a transformer classifier producing $>10{,}000\times 10{,}000$6 (Wang et al., 22 Jul 2025).

Training combines three losses. Offset regression uses Smooth L1,

$>10{,}000\times 10{,}000$7

Vertex classification uses binary cross-entropy,

$>10{,}000\times 10{,}000$8

An angle penalty encourages $>10{,}000\times 10{,}000$9 when $d_1=1<d_2<\dots<d_n$0 and $d_1=1<d_2<\dots<d_n$1 otherwise: $d_1=1<d_2<\dots<d_n$2 The total loss is

$d_1=1<d_2<\dots<d_n$3

with $d_1=1<d_2<\dots<d_n$4 and $d_1=1<d_2<\dots<d_n$5 in the reported implementation (Wang et al., 22 Jul 2025).

The PST formulation makes a specific claim about polygon vectorization: vertices are not treated as direct contour samples, but as latent salient points inferred from refined geometry and angle structure. This suggests a distinction between contour regularization and semantic vertex recovery, with the latter delegated to transformer sequence modeling.

5. Datasets, implementation, and quantitative evaluation

The reported experiments use three large-size RSI datasets. WHU-building contains 400 large tiles of size $d_1=1<d_2<\dots<d_n$6 at $d_1=1<d_2<\dots<d_n$7 m resolution, split $d_1=1<d_2<\dots<d_n$8 for train/validation/test. GLH-water contains 250 tiles of size $d_1=1<d_2<\dots<d_n$9 at $\{1,3,6\}$0 m, split $\{1,3,6\}$1. VHR-road contains 208 tiles of size $\{1,3,6\}$2 at $\{1,3,6\}$3 m, split $\{1,3,6\}$4. Multi-scale pyramids use $\{1,3,6\}$5 for buildings and $\{1,3,6\}$6 for water and roads. MCR uses $\{1,3,6\}$7, with interpolation $\{1,3,6\}$8 for buildings and $\{1,3,6\}$9 for water and roads. Optimization uses Adam with learning rate $\{1,5,10\}$0 for CAN and $\{1,5,10\}$1 for PST, with training in PyTorch on four NVIDIA A100 GPUs (Wang et al., 22 Jul 2025).

The principal quantitative comparisons are summarized below.

Dataset	Baseline	HoliTracer
WHU-building	HiSup: PoLiS 5.49, C-IoU 42.23, AP 56.77, IoU 85.79, F1 92.21	PoLiS 3.63, C-IoU 82.30, AP 61.07, IoU 91.60, F1 95.41
GLH-water	DeepSnake: PoLiS 115.51, C-IoU 43.53, AP 2.54, IoU 64.19, F1 75.42	PoLiS 81.87, C-IoU 59.24, AP 20.84, IoU 85.68, F1 91.51
VHR-road	HiSup: PoLiS 299.03, C-IoU 3.43, AP 1.51, IoU 36.40, F1 49.77	PoLiS 134.13, C-IoU 6.10, AP 1.58, IoU 46.48, F1 60.63

On WHU-building, the reported absolute gain is PoLiS $\{1,5,10\}$2, C-IoU $\{1,5,10\}$3, AP $\{1,5,10\}$4 points, and IoU $\{1,5,10\}$5 points. Visual comparisons are stated to confirm that HoliTracer produces more connected and accurate polygons for large buildings, continuous water bodies, and complex roads (Wang et al., 22 Jul 2025).

These results emphasize that the gains are not limited to mask overlap metrics such as IoU and F1. The largest reported improvements occur in polygon-sensitive measures such as C-IoU and PoLiS, which suggests that the method’s main benefit lies in vector quality rather than in segmentation quality alone.

6. Ablations, failure modes, and implications

The ablation studies isolate contributions from CAN, PST, and geometric angle modeling. For CAN on WHU-building, the best reported image pyramid is $\{1,5,10\}$6, with IoU $\{1,5,10\}$7 and F1 $\{1,5,10\}$8, compared with IoU $\{1,5,10\}$9 and F1 $512\times 512$0 for the single-scale setting $512\times 512$1. For the vectorization stage on WHU-building, CAN+TC89-KCOS yields PoLiS $512\times 512$2, C-IoU $512\times 512$3, AP $512\times 512$4, IoU $512\times 512$5, and F1 $512\times 512$6; adding Douglas–Peucker gives PoLiS $512\times 512$7, C-IoU $512\times 512$8, AP $512\times 512$9, IoU $\{I_1,I_2,\dots,I_n\}$0, and F1 $\{I_1,I_2,\dots,I_n\}$1; adding PST yields PoLiS $\{I_1,I_2,\dots,I_n\}$2, C-IoU $\{I_1,I_2,\dots,I_n\}$3, AP $\{I_1,I_2,\dots,I_n\}$4, IoU $\{I_1,I_2,\dots,I_n\}$5, and F1 $\{I_1,I_2,\dots,I_n\}$6. For angle modeling, the best result is obtained with angle features and $\{I_1,I_2,\dots,I_n\}$7, giving AP $\{I_1,I_2,\dots,I_n\}$8, C-IoU $\{I_1,I_2,\dots,I_n\}$9, PoLiS $10{,}000\times 10{,}000$00, and IoU $10{,}000\times 10{,}000$01 (Wang et al., 22 Jul 2025).

The qualitative analysis reports several characteristic effects. CAN recovers large-scene context. MCR produces clean and uniformly sampled contours. PST refines vertices to produce smooth, well-aligned polygons on irregular shapes such as water shorelines and road junctions. At the same time, the reported failure modes state that extremely dense road networks still pose connectivity challenges and that very elongated thin structures may require finer interpolation (Wang et al., 22 Jul 2025).

A common simplifying assumption in large-image mapping is that patch stitching is primarily an engineering inconvenience rather than a representational problem. The HoliTracer results argue against that assumption: the strongest gains occur precisely where global context and boundary continuity matter most. The reported future directions—real-time acceleration, integration of height or semantic attributes, multi-class simultaneous mapping including vegetation and bridges, fusion with LiDAR or multi-spectral data, and deployment in GIS and urban planning pipelines for automated map updating and change detection—suggest that the framework is intended as a general large-scene vector mapping substrate rather than a task-specific polygonizer (Wang et al., 22 Jul 2025).