Maps of buildings and roads are increasingly drawn by neural networks straight from aerial and satellite images. Many of these networks output vectors, meaning lists of corner coordinates and connected line segments, rather than colored pixel masks. The images are often tens of thousands of pixels wide, too large for a GPU, so they are cut into tiles that are processed one at a time. Every current vector extractor then makes the same mistake. When a tile edge slices through a building or a road, the model draws a closed shape or a dead end along the cut, as if the tile border were a real wall or the real end of the road. After the tiles are reassembled, one building becomes two polygons with a fake flat wall between them, and one road becomes two stubs that do not meet.
This note proposes a different contract between the tile and the full map. The per-tile model should be allowed to say "this shape continues past the border." It does this by emitting open fragments whose loose ends are marked as cut points. Each cut point carries a small learned description of what is leaving the tile. A cheap matching step on each shared tile border then joins cut points from neighboring tiles, which produces one building polygon or one continuous road. The computation involved is proportional to the length of the tile borders, not to the image area. The claim is that this cleanly separates tile-local geometry from cross-tile identity, which overlapping windows and pixel blending both fail to do. We also argue that the separation matters most for entities that no object-centered crop can contain, such as road networks.
Core Claim
Treating tile borders as places where a shape is cut open, rather than as places where it must close, should remove most seam errors in vector extraction at little extra cost. Concretely, three things should improve on entities that straddle tile borders: fewer ground-truth entities split into several predictions, fewer spurious vertices, and better road-graph connectivity across seams. The final map should also stop depending on where the tile grid happens to fall.
The advantage should be decisive for road networks on large imagery. For buildings, a simpler two-pass scheme that re-crops around objects near the seam may do nearly as well. The experiments are designed to measure exactly where that crossover lies.
Where This Sits
Recent vector extractors supply the per-tile machinery we build on, and all of them assume that the tile contains the whole object. RoIPoly crops features around each detected building and predicts an ordered list of vertices. Its authors state that separate crops cannot communicate with each other. BiSVP walks the vertices of a closed outline in both directions, so a truncated building is forced to close on itself. The Global Collinearity-aware Polygonizer turns instance masks into compact polygons through dynamic programming that rewards straight walls. Because it trusts the upstream mask, it will faithfully straighten an artificial cut line into a permanent wall. Frame field learning fits a skeleton graph to a learned field of wall directions within each patch. For roads, MaGRoad scores a candidate edge by sampling evidence along the path between two junctions. A road that exits the image leaves a path with missing evidence, and the model can then read that gap as a dead end.
Our proposal reuses these per-tile decoders and adds one idea none of them contain: a vertex type meaning "leaves the tile here," together with a way to reconcile such vertices across tiles.
Work on tile artifacts motivates the problem, but it stays in pixel space. Boundary Loss for Remote Sensing Imagery Semantic Segmentation adds a differentiable boundary-F1 term that sharpens edges of objects cut off by the tile. The Boundary-enhanced Patch-merging Transformer replaces the fixed patch grid with content-dependent tokens. The Power of Tiling for Small Object Detection documents objects that are truncated or double-counted at tile edges. All three produce masks or boxes. None of them produce shapes that know they continue elsewhere, so they constrain our design only in one respect: whatever we build must beat pixel-level fixes that are fed into a good polygonizer.
Research Program
The per-tile model is an existing vector decoder with two additions. We use a RoIPoly-style vertex decoder for buildings and a MaGRoad-style graph decoder for roads.
The first addition is a vertex-type head. It labels each predicted vertex either as a real corner or junction, or as a cut point lying on the tile border. A building clipped by the border becomes an open polyline whose two ends are cut points, instead of a closed loop.
The second addition gives each cut point a descriptor. The descriptor combines three things: the cut point's position along the border, the direction of the crossing segment, and a feature vector pooled from a thin strip of pixels along the border. The strip matters because, near a cut point, it shows roughly what the neighboring tile shows on the other side of the line.
Reconciliation happens independently on each shared border. Given the cut points from the two sides, we solve a soft assignment (Sinkhorn-normalized matching) whose cost combines three terms: the positional gap, the direction mismatch, and the descriptor distance. An extra "unmatched" option absorbs spurious cut points. Matched fragments are then joined. For buildings, a building that crosses several tiles is closed by following its matched cut points around the chain of fragments. Finally, we run the collinearity dynamic program of the Global Collinearity-aware Polygonizer only on a window around each seam. This removes the leftover cut vertex wherever the two joined segments are nearly collinear, and keeps it wherever the joint is a genuine corner. For roads, matched cut points are merged into a single graph node, and the node is deleted if it has exactly two neighbors.
Training data comes for free. We lay random tile grids over any training image that has vector labels, clip the ground-truth polygons and graphs to each tile, and mark the resulting clip points as ground-truth cut points with their true partners. The vertex-type head uses a classification loss. The matcher is trained end to end with a matching loss against the known partners. Since grids are resampled at every step, the model sees cuts in every position.
Baselines
The naive reference point is the unmodified per-tile vectorizer run on a non-overlapping grid, with outputs simply concatenated. This is how most deployed pipelines behave, and it sets the floor for the seam-specific metrics.
The first strong competitor is overlapping tiles followed by geometric union. The same per-tile vectorizer runs on tiles with 25% and 50% overlap. Predictions that overlap are merged with a polygon union for buildings, and for roads with endpoint snapping within a distance threshold. This is what a GIS practitioner would do first. We tune the overlap fraction and snap threshold on validation data and report backbone compute alongside accuracy. Beating it means better seam metrics at equal or lower backbone compute, not merely at higher compute.
The second strong competitor is raster stitching followed by vectorization. A segmentation network runs on overlapping tiles, its logits are blended with Gaussian weights, and the stitched mask is passed to the Global Collinearity-aware Polygonizer on a single large canvas. This competitor has never seen a seam in its vector stage, so it tests whether cut points buy anything over clean pixel stitching. It is trained with the boundary loss of Bokhovkin and Burnaev so that its edges are as sharp as possible.
The third competitor applies to buildings only: object-centered re-cropping. A first pass detects boxes. Any box that touches a tile border is re-cropped as a window centered on that box and re-vectorized, in the RoIPoly manner.
For reference numbers, RoIPoly reports 83.6 AP on CrowdAI with a Swin-Base backbone, and frame field learning reports a maximum tangent angle error of 31.9° on CrowdAI. Both were measured on untiled evaluation images, so they cap what the unmodified decoders achieve when there are no seams. For roads on WildRoad, MaGRoad is reported to beat SAM-Road++ on F1 and on APLS, a path-length-based measure of graph topology. However, no reported result found on Lacuna for the numeric MaGRoad F1 or APLS on WildRoad, so we will re-run it ourselves.
Two ablations isolate the key ingredient. The first replaces the learned descriptor with purely geometric matching on position and direction. The second removes the vertex-type head entirely and matches closed-polygon edges that lie on the border. If the geometric-only matcher performs as well as the full one, the contribution reduces to "emit open fragments," and we will report it that way.
Why not just…?
Why not just run overlapping tiles and take the polygon union? Two slightly misaligned predictions of the same building produce slivers, duplicate vertices, and step-shaped edges when unioned. A 50% overlap also roughly quadruples the number of backbone passes. Still, this is a real threat if a tuned overlap closes the gap. Our experiments sweep the overlap fraction and plot seam accuracy against compute, so the answer is a curve rather than an assertion.
Why not just blend logits and vectorize the stitched mask? The Global Collinearity-aware Polygonizer's authors state that polygon quality is bounded by mask quality. Pixel-based networks also round off corners, as the 31.9° versus 51.9° tangent-angle gap reported for frame fields over U-Net shows. For roads, a blended mask still has to be converted to a graph, and that conversion will be run on the whole map. The comparison in our seam-subset metrics settles whether blending already suffices.
Why not just re-crop whole objects that touch the seam? For buildings smaller than a tile, this is the most dangerous alternative, and it may well match our method. It cannot work for roads, which have no bounding box, or for buildings larger than a tile. It also adds a second pass whose cost grows with the number of buildings near seams. We will report the building-size regime in which re-cropping ties with our method and the regime in which it fails, rather than claim a uniform win.
Why not just use bigger tiles or a global-plus-local network? Bigger tiles reduce how many seams there are, but do not change what happens at a seam. Dual-branch global-local networks were outperformed by content-aware tokens in the Boundary-enhanced Patch-merging Transformer: 71.2% (GLNet) and 74.2% (ISDNet) against 77.1% mIoU on Inria Aerial. They also produce masks, not vectors. We include a tile-size sweep to show how the seam error rate scales with seam density.
Proposed Experiments
For buildings we use CrowdAI, WHU-Mix (vector labels), and SpaceNet Las Vegas. CrowdAI's 300-pixel images are too small to contain real seams, so we create seams artificially by imposing sub-tile grids of 100 and 150 pixels. Because ground truth exists for the whole image, the seam effect can be measured exactly. For roads we use WildRoad, whose 8K×4K images are the regime where tiling is unavoidable. We tile at several sizes and evaluate the reassembled graph on each full image.
We report standard metrics: AP, AP75, vertex ratio, and maximum tangent angle error for buildings, and F1 and APLS for roads. We also report them restricted to entities that intersect a grid line.
Three seam-specific measurements carry the main argument. The split rate is the fraction of ground-truth entities that are covered by more than one prediction. The spurious-vertex count is the number of predicted vertices within a few pixels of a seam that have no ground-truth corner nearby. Cross-seam connectivity for roads is the fraction of ground-truth road crossings that the output graph also connects.
The decisive check is grid-offset invariance. We shift the tile grid by random offsets and measure how much the final map changes, using polygon IoU for matched buildings and graph edit distance for roads. A method that has truly solved the seam problem should produce nearly the same map wherever the grid falls. The naive pipeline and the raster-stitching pipeline should not.
The proposal succeeds under two conditions. First, on WildRoad, cross-seam connectivity and APLS on the seam subset must rise above the overlap-and-snap baseline at no more backbone compute. Second, on buildings, split rate and spurious vertices must fall below the overlap-union and raster-stitching baselines while matching object-centered re-cropping. It fails if a tuned overlap baseline reaches the same seam metrics at comparable compute, or if the geometric-only ablation erases the gap and shows that the learned descriptor is unnecessary.
References
- RoIPoly: Vectorized building outline extraction using vertex and logit embeddings
- BISVP: Building Footprint Extraction Via Bidirectional Serialized Vertex Prediction
- Global Collinearity-Aware Polygonizer for Polygonal Building Mapping in Remote Sensing
- Polygonal Building Extraction by Frame Field Learning
- Beyond Endpoints: Path-Centric Reasoning for Vectorized Off-Road Network Extraction
- Boundary Loss for Remote Sensing Imagery Semantic Segmentation
- Ultra-High Resolution Segmentation via Boundary-Enhanced Patch-Merging Transformer
- The Power of Tiling for Small Object Detection