Classifying linear infrastructure elements using a graph neural network
In example embodiments, improved techniques are provided for classifying elements of an infrastructure model that represents linear infrastructure (e.g., roads). The techniques may extract a set of cross sections perpendicular to a centerline of the linear infrastructure from the infrastructure model, generate a graph representation of each cross section to produce a set of graphs having nodes that represent elements and edges that represent contextual relationships, provide the set of graphs to a trained graph neural network (GNN) model, and produce therefrom class predictions for the elements. The class predictions may include one or more predicted classes for each element with a respective confidence. A best predicted class for each element may be selected and assigned to the element, thereby creating a new version of the infrastructure model. For elements that extend through multiple cross sections, the selection may involve aggregating predicted classes originating from the different graphs.
1 . A method for classifying elements in an infrastructure model that represents linear infrastructure using a trained graph neural network (GNN) model, comprising:
accessing, by software executing on one or more computing devices, the infrastructure model that represents linear infrastructure;
extracting, by the software, a set of cross sections perpendicular to a centerline of the linear infrastructure from the infrastructure model, wherein the extracting cuts each cross section from a representation of the infrastructure model at an interval;
generating, by the software, a graph representation of each cross section based on information in the infrastructure model to produce a set of graphs having nodes that represent elements in the cross section and edges that represent contextual relationships between the elements;
predicting, by the software using the trained GNN model, classes based on the set of graphs to produce class predictions for the elements represented by the nodes;
selecting a predicted class of each of the elements from the class predictions; and
assigning, by the software, predicted classes to respective elements to create a new version of the infrastructure model that is maintained in memory or storage of the one or more computing devices.
2 . The method of claim 1 , wherein each class prediction includes one or more predicted classes having a respective confidence.
3 . The method of claim 2 , wherein the selecting further comprises:
for one or more elements that extend through multiple cross sections and thereby are represented by nodes included in multiple graphs, aggregating predicted classes originating from different graphs.
4 . The method of claim 3 , wherein the aggregating further comprises:
taking a predicted class from the class prediction originating from each graph based on the respective confidence of each predicted class, to create a set of predicted classes for the element; and
selecting a predicted class from among the set of predicted classes.
5 . The method of claim 2 , wherein the selecting further comprises:
for one or more elements that extend through a single cross section and thereby are represented by a node in a single graph, choosing the predicted class based on the respective confidence of each predicted class.
6 . The method of claim 1 , wherein the nodes further represent properties of the elements in the cross section, and at least one property is a geometric property.
7 . The method of claim 6 , wherein the geometric property is selected from the group consisting of width, height, area, slope angle, and number of points.
8 . The method of claim 1 , wherein the contextual relationships represented by the edges include geometric information.
9 . The method of claim 8 , wherein the geometric information includes at least one item selected from the group consisting of distance, angle, and number of overlapping points.
10 . The method of claim 1 , further comprising:
displaying, in a user interface of the software, a visualization of the new version of the infrastructure model with indications of the assigned classes; and
prompting a user to manually review the assigned classes and to correct any errors in the user interface.
11 . The method of claim 1 , wherein the linear infrastructure is a road, and the predicted class for one or more of the elements is a roadway class, a traffic lane class, a shoulder class, a roadside class, a curb class, a sidewalk class, or a central reserve class.
12 . A computing device, comprising:
a processor; and
a memory coupled to the processor and configured to store a software application configured to classify elements in an infrastructure model that represents linear infrastructure, the software application including:
a cross section extraction process configured to extract a set of cross sections perpendicular to a centerline of the linear infrastructure from the infrastructure model by cutting cross sections from a representation of the infrastructure model at an interval,
a graph generation process configured to generate a graph representation of each cross section based on information in the infrastructure model to produce a set of graphs having nodes that represent elements in the cross section and edges that represent contextual relationships between the elements,
a trained graph neural network (GNN) model configured to predict classes based on the set of graphs to produce class predictions for the elements represented by the nodes, and
a selection and aggregation process configured to, for one or more elements that extend through multiple cross sections and thereby are represented by nodes included in multiple graphs, aggregate class predictions originating from different graphs.
13 . The computing device of claim 12 , wherein the linear infrastructure is a road.
14 . A non-transitory electronic-device readable media having instructions stored thereon that when executed on one or more processors of one or more computing devices are operable to:
access an infrastructure model that represents linear infrastructure;
extract a set of cross sections perpendicular to a centerline of the linear infrastructure from the infrastructure model by cutting cross sections from a representation of the infrastructure model at an interval;
generate a graph representation of each cross section based on information in the infrastructure model to produce a set of graphs having nodes that represent elements in the cross section and edges that represent contextual relationships between the elements;
predict, using a trained graph neural network (GNN) model, classes based on the set of graphs to produce class predictions for the elements represented by the nodes,
for one or more elements that extend through multiple cross sections and thereby are represented by nodes included in multiple graphs, aggregate class predictions originating from different graphs; and
assign a predicted class from the aggregated class predictions to respective elements to create a new version of the infrastructure model.
15 . The non-transitory electronic-device readable media of claim 14 , wherein the instructions when executed are further operable to:
for the one or more elements that extend through multiple cross sections, take a predicted class from the class prediction originating from each graph based on the respective confidence of each predicted class, to create a set of predicted classes for the element; and select a predicted class from among the set of predicted classes.
16 . The non-transitory electronic-device readable media of claim 14 , wherein the instructions when executed are further operable to:
for one or more elements that extend through a single cross section and thereby are represented by a node in a single graph, choose a predicted class based on the respective confidence of each predicted class.
17 . The non-transitory electronic-device readable media of claim 14 , wherein the nodes further represent properties of the elements in the cross section, and at least one property is selected from the group consisting of width, height, area, slope angle, and number of points.
18 . The non-transitory electronic-device readable media of claim 14 , wherein the contextual relationships represented by the edges include at least one item of geometric information selected from the group consisting of distance, angle, and number of overlapping points.
19 . The non-transitory electronic-device readable media of claim 14 , wherein the instructions when executed are further operable to:
display a visualization of the new version of the infrastructure model with indications of the assigned classes; and
prompt a user to manually review the assigned classes and to correct any errors in the user interface.
20 . The non-transitory electronic-device readable media of claim 14 , wherein the linear infrastructure is a road and the predicted class for one or more of the elements is a roadway class, a traffic lane class, a shoulder class, a roadside class, a curb class, a sidewalk class, or a central reserve class.