# MatchGeo Annotation Protocol
# Local Feature Matching Keypoint Annotation on Digital Elevation Models

> **Version**: 1.1  
> **Date**: 2026-07-24  
> **Authors**: Sabrina Correa (UFV)  
> **Dataset**: MatchGeo DEM v1.1 (DOI: 10.5281/zenodo.21229785)  
> **Standard**: Aligned with OGC TrainingDML-AI (23-008r3 / 24-006r1)  
> **License**: CC BY 4.0

---

## Table of Contents

1. [Purpose and Scope](#1-purpose-and-scope)
2. [Prerequisites](#2-prerequisites)
3. [Annotation Class](#3-annotation-class)
4. [Annotation Workflow](#4-annotation-workflow)
5. [Annotation Rules](#5-annotation-rules)
6. [Quality Control](#6-quality-control)
7. [File Format and Metadata](#7-file-format-and-metadata)
8. [Common Pitfalls and Edge Cases](#8-common-pitfalls-and-edge-cases)
9. [Appendix A: Visual Example](#appendix-a-visual-example)
10. [Appendix B: Glossary](#appendix-b-glossary)

---

## 1. Purpose and Scope

### 1.1 Objective

This protocol defines the procedure for annotating salient elevation keypoints on 256x256 pixel Digital Elevation Model (DEM) tiles. The annotated keypoints serve as ground truth for training and evaluating **local feature matching algorithms** — computer vision methods that detect and describe distinctive points in images (or elevation surfaces) to establish correspondences across different views, scales, or domains.

### 1.2 What to Annotate

Annotators must identify **morphologically salient points** on the DEM surface that would be reliably detected by feature descriptors such as SIFT, SURF, ORB, or learned descriptors (e.g., SuperPoint, R2D2). These points must be:

- **Locally distinctive**: The point must stand out from its immediate neighborhood.
- **Geometrically stable**: The point must remain detectable under moderate changes in resolution, noise, or viewing angle.
- **Repeatable**: The point should be identifiable by another annotator (or the same annotator on a different day) with high spatial consistency.

In **urban areas** (e.g., São Paulo, Bonn), the primary salient features are **building corners** — the intersection points of building roof edges where elevation changes abruptly.

### 1.3 What NOT to Annotate

Do **not** annotate:

- Points in flat or uniformly sloped areas with no local distinctiveness.
- Points at tile boundaries (within 5 pixels of the edge).
- Points in areas with obvious data artifacts (striping, voids, interpolation errors).
- Points in water bodies (lakes, rivers) unless the shoreline intersection is clearly salient.
- Points on roads, sidewalks, or other flat man-made surfaces unless they form a clear corner or edge intersection.
- **Points where the annotator is not fully confident about the salience.**

### 1.4 Scope of Application

This protocol applies, but not limited to:

- **GER_BN** (Bonn, Germany) — primary annotated city
- **BRA_SP** (São Paulo, Brazil) — secondary annotated city
- Any future cities added to the MatchGeo dataset with manual annotations

---

## 2. Prerequisites

### 2.1 Required Software

- **QGIS** ≥ 3.28 (Long Term Release)
- **Python** ≥ 3.9 with `geopandas`, `rasterio`, `shapely`

### 2.2 Required Data

For each tile to annotate:

| File | Purpose |
|---|---|
| `{CITY_CODE}_{row}_{col}.tif` | 256x256 pixel DEM tile (Float32, NoData=-9999) |


### 2.3 Annotator Qualifications

Annotators should have:

- Basic familiarity with DEM visualization techniques (hillshade, slope).
- Experience with QGIS or equivalent GIS software.
- Understanding of what makes a point "salient" for computer vision feature detectors.

---

## 3. Annotation Class

All annotated keypoints belong to a **single class**:

### `general_interest_point` — Salient Elevation Keypoint

> **Definition**: A point on the DEM surface that is locally distinctive and suitable for local feature matching. In urban areas, this is typically a **building corner** (the intersection of roof edges). In natural terrain, this is any point with high local curvature or distinctiveness that would produce a strong response from a feature detector.

**Rationale for single class**: The MatchGeo dataset is designed for **local feature matching**, not semantic segmentation or landform classification. The task requires correspondences between geometrically similar points, not semantic labels. A single class simplifies annotation while preserving the geometric information needed for feature matching research.

---

## 4. Annotation Workflow

### 4.1 Pre-Annotation Setup

1. **Load the tile** in QGIS.
2. **Verify CRS**: Ensure the project CRS matches the tile CRS (e.g., `EPSG:25832` for Bonn, `EPSG:31983` for São Paulo).
3. **Chenge visualization to hillshade** as a visual reference (optional but strongly recommended).
4. **Set zoom level**: Work at a scale where individual pixels are visible but the full tile context is retained.
5. **Check for artifacts**: Scan the tile for data quality issues (voids, striping, noise). Skip tiles with >20% artifact coverage.

### 4.2 Annotation Procedure

For each tile, follow this sequence:

#### Step 1: Global Scan
- Pan across the entire tile at moderate zoom.
- Identify the dominant features: buildings, vegetation, roads, natural terrain.
- Note any artifact regions to avoid.

#### Step 2: Systematic Pass
- Scan the tile in a systematic pattern (left-to-right, top-to-bottom).
- For each candidate point, ask: **"Would a feature detector (SIFT, Harris, SuperPoint) reliably find this point?"**
- Place the keypoint at the **precise pixel** of maximum distinctiveness.
- **Only annotate if you are fully confident** in the point's salience. When in doubt, skip.

#### Step 3: Verification Pass
- Toggle hillshade on/off to verify distinctiveness.
- Check that no keypoints are within **5 pixels** of the tile edge.
- Verify that keypoints are not clustered — minimum spacing of **5 pixels** between any two keypoints.
- Review each point: **"If I saw this tile again tomorrow, would I place the point at the same pixel?"**

#### Step 4: Save
- Save the QGIS project with the annotation shapefile layer.
- Export the annotation layer as a **shapefile** or **GeoPackage**.

### 4.3 Post-Processing

After annotation, run the MatchGeo conversion script to transform the shapefile into the standardized JSON annotation format. The conversion file is in:

```
scripts/convert_shp2anno.py
```

The script will:
1. Read the shapefile points.
2. Convert geographic coordinates to pixel coordinates (0–255).
3. Extract elevation values from the DEM tile.
4. Generate the TDML-compliant JSON file.

---

## 5. Annotation Rules

### 5.1 Minimum Feature Distinctiveness

All annotated features must be resolvable at the tile resolution. The point must have:

- **Local curvature**: The point must be a local extremum (maximum or minimum) in at least one direction, OR a point of high curvature (corner, junction).
- **Minimum neighborhood**: The distinctiveness must be visible in a 5×5 pixel neighborhood.
- **No ambiguity**: The point must have a single, clear location. If the "best" pixel is ambiguous across a 3×3 region, do not annotate.

### 5.2 Edge Proximity

- **Default**: No keypoints within 5 pixels of the tile boundary.
- **Hard limit**: Never annotate within 2 pixels of the boundary.
- **Rationale**: Points near boundaries may be truncated or padded differently during model training, leading to inconsistent feature descriptors.

### 5.3 Spacing

- Minimum distance between any two keypoints: **5 pixels**.
- **Rationale**: Feature descriptors typically operate on patches of 8×8 to 32×32 pixels. Points closer than 5 pixels would have overlapping descriptor windows, reducing independence.

### 5.4 Urban Areas: Building Corners

In urban DEMs (Bonn, São Paulo), the vast majority of salient points are **building corners**:

- **What is a building corner?** The intersection of two roof edges, where the elevation surface changes direction abruptly.
- **How to identify**: On hillshade, building corners appear as sharp, dark-light transitions at the intersection of two edges.
- **Placement**: Place the point at the **intersection pixel** — the exact corner where the two edges meet.
- **Avoid**: Points along straight roof edges (not corners), points on flat roof surfaces, points on building shadows (unless the shadow edge coincides with a real corner).

### 5.5 Natural Terrain

In natural terrain (if annotated in future cities), salient points include:

- **Peaks and summits**: Local maxima with clear prominence.
- **Saddles and cols**: Local minima along ridge lines.
- **Ridge/valley junctions**: Confluences of multiple ridge or valley lines.
- **Cliff corners**: Abrupt changes in slope direction.

When in doubt, apply the **feature detector test**: *"Would I select this point tomorrow?"*

### 5.6 Multi-Scale Considerations

Annotators should consider how the feature would appear at different resolutions:

- A feature visible at 1.0 m resolution should still be detectable (though blurred) at 2.0 m.
- Do not annotate features that are **resolution-dependent artifacts** (e.g., single-pixel noise spikes).
- When in doubt, check the source data quality or hillshade at multiple sun angles.

---

## 6. Quality Control

### 6.1 Self-Review Checklist

Before submitting annotations, the annotator must complete this checklist:

- [ ] All keypoints are within the valid tile area (≥5 px from edge).
- [ ] No two keypoints are within 5 pixels of each other.
- [ ] All keypoints are on clearly salient features (building corners or natural morphological features).
- [ ] No keypoints are in obvious artifact areas (voids, striping).
- [ ] All keypoints would be repeatable by the same annotator on a different day.
- [ ] The annotation shapefile has been exported and converted to JSON.
- [ ] The JSON file validates against the MatchGeo Annotation Schema.


### 6.2 Confidence Policy

All annotated keypoints are assumed to have **high confidence** (≥0.8). The annotator is instructed to:

- **Only annotate points they are fully confident about.**
- **Skip ambiguous points.** It is better to have fewer, high-quality annotations than many uncertain ones.
- **Do not guess.** If a point might be salient but the annotator is not sure, skip it.

For retroactive application to existing annotations (GER_BN, BRA_SP), all keypoints are assigned a default confidence of **0.8**.

---

## 7. File Format and Metadata

### 7.1 QGIS Annotation Layer

During annotation, points are stored in a QGIS vector layer (shapefile or GeoPackage):

| Field | Type | Description |
|---|---|---|
| `id` | Integer | Unique keypoint ID within the tile |
| `class` | String | Fixed `"general_interest_point"` |
| `confidence` | Float | Annotator confidence (0.0–1.0); default 0.8 |

### 7.2 Annotation File Naming

```
{CITY_CODE}_{row}_{col}_annotation.json

Example: GER_BN_001_012_annotation.json
```

### 7.3 JSON Structure

Each annotation file conforms to the MatchGeo Keypoint Annotation Schema:

```json
{
  "type": "AI_ObjectLabel",
  "id": "GER_BN_001_012_annotation",
  "tileId": "GER_BN_001_012",
  "cityCode": "GER_BN",
  "crs": "EPSG:25832",
  "resolution": 1.0,
  "tileExtent": {
    "westBound": 369500.0,
    "eastBound": 369833.0,
    "southBound": 5621500.0,
    "northBound": 5621833.0
  },
  "keypoints": [
    {
      "id": "kp_001",
      "pixelX": 312,
      "pixelY": 282,
      "geoX": 369812.0,
      "geoY": 5621782.0,
      "class": "general_interest_point",
      "elevation": 158.4,
      "confidence": 0.8
    }
  ],
  "labeling": {
    "labelingMethod": "manual",
    "annotatorId": "annotator_ufv_01",
    "annotationTime": "2025-11-14T14:32:00Z",
    "toolVersion": "qgis-shapefile-conversion",
    "labelingProtocol": "https://github.com/paeslemesa/matchgeodem/blob/main/docs/annotation_protocol.md"
  },
  "quality": {
    "sourceDataQuality": "Airborne LiDAR, ground classified, vertical accuracy <0.15m RMSE"
  }
}
```

### 7.4 Required vs. Optional Fields

| Field | Required | Description |
|---|---|---|
| `type` | ✅ | Fixed `"AI_ObjectLabel"` |
| `id` | ✅ | Unique annotation file ID |
| `tileId` | ✅ | Corresponding DEM tile ID |
| `cityCode` | ✅ | City code (e.g., `"GER_BN"`, `"BRA_SP"`) |
| `crs` | ✅ | EPSG code |
| `resolution` | ✅ | GSD in meters |
| `tileExtent` | ❌ | Bounding box in projected coords |
| `keypoints` | ✅ | Array of keypoint objects |
| `keypoints[].id` | ✅ | Unique keypoint ID |
| `keypoints[].pixelX` | ✅ | X in pixel space (0–332) |
| `keypoints[].pixelY` | ✅ | Y in pixel space (0–332) |
| `keypoints[].geoX` | ❌ | X in projected CRS |
| `keypoints[].geoY` | ❌ | Y in projected CRS |
| `keypoints[].class` | ✅ | Fixed `"general_interest_point"` |
| `keypoints[].elevation` | ✅ | Elevation at keypoint (m) |
| `keypoints[].confidence` | ❌ | Annotator confidence (0–1); default 0.8 |
| `labeling` | ✅ | Provenance object |
| `labeling.labelingMethod` | ✅ | Fixed `"manual"` |
| `labeling.annotatorId` | ✅ | Anonymous annotator ID |
| `labeling.annotationTime` | ✅ | ISO 8601 timestamp |
| `labeling.toolVersion` | ❌ | `"qgis-shapefile-conversion"` |
| `labeling.labelingProtocol` | ❌ | URL to this protocol |
| `quality` | ❌ | Quality metrics object |

---

## 8. Common Pitfalls and Edge Cases

### 8.1 Building Corners vs. Roof Edges

**Problem**: A long, straight roof edge has many points that look somewhat distinctive.

**Solution**: Only annotate **corners** — points where two edges intersect. A point along a straight edge is not locally distinctive in all directions and would not produce a strong corner detector response.

### 8.2 Flat Roofs

**Problem**: Large flat-roofed buildings have no corners visible in the DEM.

**Solution**: Do not annotate flat roof surfaces. If the building has a parapet or edge wall that creates a corner, annotate that corner.

### 8.3 Vegetation

**Problem**: Trees and vegetation create noisy, irregular elevation surfaces.

**Solution**: In urban areas, focus on building corners. In natural terrain, avoid annotating on dense vegetation unless a clear morphological feature (e.g., a forest clearing corner) is visible.

### 8.4 Data Artifacts

**Problem**: Striping, voids, or interpolation artifacts create spurious elevation features.

**Solution**:
- Learn to recognize common artifacts:
  - **Striping**: Parallel lines of alternating high/low elevation (common in satellite DEMs).
  - **Voids**: Areas with NoData (-9999) or interpolated fill values.
  - **Noise**: Single-pixel spikes or pits with no surrounding morphological context.
- Cross-reference with hillshade.
- Skip artifact-affected tiles if >20% of the tile is affected.

### 8.5 Partial Tiles

**Problem**: Tiles at the city boundary may be partially empty (NoData).

**Solution**: Annotate only the valid data region. Ensure all keypoints are on valid elevation pixels (not NoData).

### 8.6 Multi-Story Buildings

**Problem**: Buildings of different heights create complex roof patterns.

**Solution**: Annotate the corners of the **roof footprint** — the outermost edges visible in the DEM. Internal roof structures (skylights, HVAC units) are typically too small to be reliable at 1.0 m resolution.

---


*Protocol version 1.1 — MatchGeo DEM Annotation. For questions, contact sabrina.correa@ufv.br*