Introduction
Building3D consists of building point clouds, roof point clouds, mesh models and wireframe models. The largest city Tallinn includes about 37,000 labelled building objects. In addition, we also provide excutable programs to calculate 3D mesh IoU and Root Mean Square Error (RMSE).
Dataset Download
Please enjoy this demo dataset we have provided. Click here to download.
- Entry-level Dataset Roof Point Clouds & Wireframe [6,000 buildings / 260M]
- Tallinn City: Building Point Clouds [47,000 buildings / 5.6G] Roof Point Clouds & Wireframe [36,000 buildings / 1.7G] Mesh (NOT recommend) [47,000 buildings / 107M]
- Other Cities: Coming soon
Data Type
Building Point Clouds
Each building point clouds are stored in XYZ format including XYZ coordinates, RGB colour, near-infrared information, intensity and reflectance.
Roof Point Clouds
For 3D roof reconstruction, it doesn't involve facade point clouds. Thus, the roof point clouds only retain all the points representing roof structure.
Mesh
Building mesh models are created from aerial LiDAR point clouds and building footprints by using the Terrasolid software, and then modified by hand.
Wireframe
Wireframe models are a very simple 3D representation. It consists of vertexes and edges.
Roof Type
Building3D dataset has over 60 roof types. About ten frequently encountered roof types are shown below.
Hexagonal Gazebo
Hip Roof
M Shaped
Butterfly
Flat
Cross Gable
Cross Hipped
Intersecting Hip
Skillion and Lean to
Flat
Dutch Roof
Gable Roof
Data Preview
The video shows the type of data in a certain area of the display.
Evaluation Metrics
Average corner offset (ACO) is the average offsets between predicted corners and ground-truth corners. The smaller offsets indicate better quality of generated models.
Corner precision (CP), edge precision (EP), corner recall (CR), and edge recall (ER) are calculated through confusion matrix to evaluate the accuracy of corner and edge classification. larger CP and EP values indicate more precise classification of corners and edges, while larger CR and ER values indicate lower rates of missing detection.
3D Mesh IoU is a metric for evaluation of the fit between generated mesh models and ground truth mesh models. We develop a numerical solution to use mesh models for 3D IoU to represent fitting errors.
Root mean square (RMS) distance is a metric for evaluation of the fit between input roof point clouds and generated mesh models.
Support Material
