Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →DINO-EdgeQuery is an edge-first computer-vision pipeline that turns overhead images into editable building polygons. Instead of tracing a raster mask and simplifying its outline, it predicts wall-like edges, estimates how those edges connect, intersects their supporting lines to obtain corners, and assembles each closed outline with deterministic geometry checks. With georeferenced input, the resulting coordinates can be exported as GeoJSON for QGIS or WebGIS workflows.
The method combines a frozen DINOv3 image backbone and adapter neck with an instance decoder and an ROI-conditioned EdgeQuery decoder. Its reported validation used 0.2 m aerial photography from the WHU Building Dataset in 512 × 512-pixel tiles, so the published numbers should not be treated as a blanket result for every satellite sensor or geography.
As an Amazon Associate I earn from qualifying purchases.
Why an edge-first pipeline is useful
Building extraction has two different jobs that are often conflated. The first is recognition: deciding which pixels belong to a building. The second is cartography: converting those pixels into a compact, editable polygon with sensible corners, a closed boundary and a known coordinate reference system.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A segmentation mask is good at the first job, but its pixel boundary is not automatically a useful GIS feature. Converting a mask to a contour can create stair-stepped edges, noisy corners, self-crossings or duplicate outlines where buildings touch. Smoothing and polygon simplification can reduce noise, but those operations do not know which lines represent walls or how separate edges should be connected.
#1 Best Overall
DINO-EdgeQuery addresses that raster-to-vector gap by making edges and their connections explicit. The neural network learns which candidate edges are active and how they relate; deterministic geometry then turns those predictions into polygons. This division keeps the learned part focused on visual recognition while making assembly and validity checks inspectable and repeatable.
What DINO-EdgeQuery predicts
Backbone and instance recognition
A frozen DINOv3 backbone and adapter neck produce image features. An instance decoder uses those features to propose building-aware classes, boxes and masks. The instance branch supplies the candidate building regions in which the edge decoder will operate.
ROI-conditioned EdgeQuery decoding
For each region of interest (ROI), the EdgeQuery decoder predicts several properties of potential boundary segments:
- Edge activity: whether a candidate segment corresponds to a building boundary.
- Edge center: where that segment is located in the ROI.
- Direction: the orientation of the wall-like segment.
- Length: the extent of the segment.
- Directed successor relations: which edge should follow another when walking around the building.
The successor relation is important because a set of correctly detected lines is not enough. A polygon requires an ordered cycle. EdgeQuery supplies a learned connectivity signal, while the subsequent ordering and geometry are deterministic.
From supporting lines to corners
Each selected edge defines a supporting line. Polygon vertices are calculated at intersections between neighboring supporting lines, rather than being emitted as a free-standing list of pixel coordinates. This tends to produce corners that are consistent with the predicted wall directions and lengths. The assembled cycle is then checked for geometric validity before it is retained.
How the full pipeline works
- Read the overhead image and georeferencing. The image may be processed in pixel coordinates first, but georeferenced inputs carry the transform and coordinate reference system needed for GIS output.
- Extract visual features. The frozen DINOv3 backbone and adapter neck create feature maps that describe roofs, boundaries and surrounding context.
- Detect building instances. The instance decoder predicts classes, bounding boxes, masks and boundary-aware signals. These predictions define the ROIs supplied to EdgeQuery.
- Tokenize each ROI. The ROI tokenizer converts the instance region and its visual features into inputs for the edge decoder.
- Predict edge primitives and links. EdgeQuery estimates activity, center, direction, length and directed successor relations for candidate boundary edges.
- Select and order edges. Confidence thresholds determine which edges are active. Successor predictions provide an order that can be formed into a closed cycle.
- Compute vertices. Intersections of adjacent supporting lines become polygon corners.
- Apply deterministic checks. Polygon assembly, validity checks and duplicate suppression are performed as explicit geometry operations rather than learned neural layers.
- Transform and export coordinates. For georeferenced imagery, pixel coordinates are converted through the source transform and coordinate reference system before export as GeoJSON or loading into a GIS application.
Training and inference are deliberately staged
G1: establish reliable building instances
In the first training stage, called G1, the instance branch learns building classes, boxes, masks and boundary-aware signals. The backbone and EdgeQuery components remain frozen. This stage gives the system a stable source of building ROIs before edge learning begins.
G2: learn edges under stable ROI conditions
In G2, the instance branch is frozen. The ROI tokenizer, EdgeQuery decoder and edge and successor heads are trained using stable mask and box conditions. Separating the stages prevents a moving instance detector from destabilizing the edge and connectivity objectives.
Recommended Free Tools
Unified inference does not use oracle masks
After training, the weights from the stages are merged for unified inference. At run time the system starts with an image and its own predicted boxes and masks; it does not require ground-truth masks or manually supplied ROIs. This matters when estimating real deployment behavior, because any instance-recognition error is present in the edge stage’s input.
P1 and P2: two inference-time quality profiles
The project documentation describes P1 and P2 as quality profiles applied during inference, not as separate neural networks. Both profiles use checks that can reject an implausible polygon after assembly.
P1 checks
P1 applies mask-consistency, coverage, spill or outside-ratio, area and geometric-validity checks. These tests compare the generated polygon with the predicted building region and remove shapes that are too small, extend too far outside the ROI or fail geometric validity.
P2 adds containment-based suppression
P2 performs the P1 checks and adds containment-based suppression for nested or duplicate polygons. When one candidate is contained within another in a way that indicates duplication, the extra output can be removed. P2 therefore reduces the number of reported polygons without changing the EdgeQuery network itself.
Reported validation results
The principal comparison was run on nonempty validation tiles from a DINO-EdgeQuery-ready version of the WHU Building Dataset. The imagery is 0.2 m aerial photography divided into 512 × 512-pixel tiles. Orbitalnet reported the following values in 2026:
| Measure | P1 | P2 | What changes |
|---|---|---|---|
| Predicted polygons | 21,554 | 19,203 | P2 emits fewer polygons after containment suppression. |
| Union IoU | 0.7107 | 0.7064 | P2 is slightly lower on this overlap measure. |
| Precision@75 | 0.5486 | 0.5924 | P2 has higher strict-match precision. |
| Recall@75 | 0.6459 | 0.6213 | P2 gives up some strict-match recall. |
| FP / GT | 0.2997 | 0.2295 | P2 produces fewer false positives per ground-truth building. |
| Self-intersections | 0 | 0 | No self-intersections were recorded under these validation conditions. |
| Triangle collapses | 0 | 0 | No triangle collapses were recorded under these validation conditions. |
The practical trade-off is straightforward: choose P2 when cleaner output and stricter matching matter more than retaining every possible building candidate. P1 retains more predictions and higher recall on this test, while P2 raises Precision@75 and lowers FP/GT. The zero counts for self-intersections and triangle collapses describe this validation setup; they are not a guarantee of perfect positional accuracy or complete detection in every image.
Empty-tile behavior is reported separately in the project materials. P2 reduces predictions on background-only tiles, so false positives on empty tiles should be evaluated separately from polygon quality on tiles that contain buildings.
Getting GIS-ready output
“GIS-ready” has a specific meaning here. It means that a polygon has coordinates tied to the source image’s georeferencing and a geometry that can be validated and edited in a GIS. It does not mean that the feature is automatically aligned to OpenStreetMap or another external map database.
- Start with georeferenced imagery. Preserve the source affine transform and coordinate reference system (CRS) when tiling or preprocessing.
- Run unified inference. Let the instance branch generate boxes and masks, then let EdgeQuery create and connect edges.
- Keep the selected quality profile. Use P1 or P2 according to whether recall or stricter duplicate suppression is the priority.
- Convert pixels to map coordinates. Apply the source georeferencing transform to every polygon vertex and retain the CRS definition.
- Export GeoJSON. Store each building polygon as a feature with its coordinates and any desired confidence or identifier attributes.
- Load and inspect in QGIS or WebGIS. Check layer CRS, overlay alignment, polygon validity and edge placement against the source image before editing or analysis.
If the input image has no reliable georeferencing, the output can still be a valid image-coordinate polygon, but it cannot be treated as a map-coordinate layer without a separate registration step.
Was it tested on satellite imagery?
The method’s title refers to satellite imagery, but the main quantitative results available for comparison come from WHU aerial imagery at 0.2 m resolution. That distinction matters: spatial resolution, viewing angle, seasonal appearance, roof materials, shadows and sensor characteristics can all change edge visibility and instance-recognition quality.
The WHU numbers therefore support claims about the stated aerial-tile setup, not automatic deployment across commercial satellite products, lower-resolution scenes or different countries. A satellite deployment should be validated with imagery from the target sensor and geography, using the same polygon metrics and empty-tile tests rather than assuming aerial performance transfers unchanged.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What can make the output fail
Missed instances cannot be recovered downstream
EdgeQuery operates on ROIs supplied by the instance branch. If the upstream recognizer misses a building, there is no ROI in which EdgeQuery can discover it. Improving edge decoding cannot compensate for a systematically weak detector in dense neighborhoods, heavy shadow or unusual roof materials.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchThresholds change the result
Confidence thresholds, non-maximum suppression (NMS), edge-selection thresholds and quality-gate settings all affect how many polygons survive. A configuration tuned for maximum recall may retain more weak or duplicate candidates; a configuration tuned for precision may discard small or partially occluded buildings.
One exterior loop is the current shape assumption
The current representation assumes one exterior closed loop per building. Courtyards and other holes, multipart buildings and more adaptive quality checks are identified as extension topics. A building with an internal courtyard may therefore need post-processing or a representation that supports interior rings.
Valid topology is not the same as exact alignment
Deterministic validity checks can prevent self-crossing rings and collapsed triangles, but they cannot guarantee that every corner is located exactly on the roof edge. Registration errors, blurred boundaries, occlusion and a wrong instance mask can still produce a geometrically valid polygon in the wrong place.
How to evaluate a deployment responsibly
- Match the evaluation imagery to the intended sensor, resolution, geography and season.
- Report both nonempty-tile quality and empty-tile false-positive behavior.
- Track strict-match precision and recall together; P2’s precision gain comes with a recall reduction in the reported validation.
- Inspect polygon validity separately from positional accuracy and missed-building rates.
- Test threshold and NMS settings on a held-out set rather than carrying defaults unchanged into a new region.
- Check how the workflow handles touching buildings, tiny structures, shadows and buildings with courtyards.
- Verify the CRS and overlay in a GIS before using areas, distances or change-detection results.
Where DINO-EdgeQuery fits among polygonization approaches
Mask-contour tracing starts with a raster boundary and usually applies contour extraction and simplification. Vertex-centric polygon models predict corner points directly and then infer or enforce an ordering. DINO-EdgeQuery instead treats wall-like edges and directed connections as the primary learned representation, deriving corners from intersections of supporting lines.
This design gives the method a clear separation between recognition and geometry: neural predictions identify plausible building edges, while deterministic routines assemble cycles, test validity and suppress duplicates. The benefit is an explicit, inspectable route from edge evidence to a GIS ring. The cost is that the system remains dependent on the instance branch and on thresholds that decide which edges and polygons survive.
Bottom line for practitioners
DINO-EdgeQuery is best understood as a learned edge-and-connectivity layer placed between building detection and deterministic GIS geometry. On the reported WHU aerial validation, P2 produced fewer polygons, higher Precision@75 and lower FP/GT than P1, while sacrificing some recall and Union IoU. Its GeoJSON path is practical when imagery is georeferenced, but satellite-wide claims, buildings with holes or multiple parts, and performance outside the tested aerial conditions still require separate validation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
