Conflation of geospatial points of interest and ground-level imagery
A prediction system harvests geo-tagged ground-level images through one or more algorithms. The system receives point of interest data representing structures or events and tags the geo-tagged ground-level images with a probability describing a classification. The system tags point of interest data with a hierarchical genre classification and encodes the tagged geo-tagged ground-level images as vectors to form nodes and edges in a proximity graph. The system encodes tagged points of interest data as similarity vectors to render more nodes and more edges on the proximity graph associated with the tagged geo-tagged ground-level images nodes by calculated semantic distances. The system splits the proximity graph into a training subgraph and a testing subgraph and trains a neural network by aggregating and sampling information from neighboring nodes within the training subgraph graph and validates through the testing subgraph. Training ends when a loss measurement is below a threshold.
1. A non-transitory machine-readable medium encoded with machine-executable instructions, wherein execution of the machine-executable instructions is for:
harvesting geo-tagged ground-level images through a scene detection algorithm executed by an image classifier;
receiving point of interest data representing structures or events through a transceiver;
tagging each of the geo-tagged ground-level images with a probability object describing a tag classification;
tagging each of the point of interest data with a hierarchical level genre classification object;
encoding tagged geo-tagged ground-level images as hot encoded vectors to render a first plurality of nodes and a first edge on a proximity graph;
encoding tagged point of interest data as a similarity vector to render a second plurality of nodes and a second edge on the proximity graph associated with the first node by a calculated semantic distance measurement;
splitting the proximity graph into a training subgraph and a testing subgraph;
initiating a training session of a neural network by processing the training subgraph during a training session that includes aggregating information at a selected node within the training subgraph by sampling information from a neighboring node within the training subgraph;
terminating the training session when a binary cross-entropy loss measurement based on the training subgraph falls below a predetermined threshold; and
initiating an evaluation of the neural network by processing the testing subgraph.
2. The non-transitory machine-readable medium of claim 1 , where the calculated semantic distance measurement comprises a calculated Euclidean distance measurement.
3. The non-transitory machine-readable medium of claim 1 , where the proximity graph is constructed from a plurality of dot products.
4. The non-transitory machine-readable medium of claim 1 , where the binary cross-entropy loss measurement comprises an average loss measurement based on a predicted edge value and a calculated edge value.
5. The non-transitory machine-readable medium of claim 4 , where the binary cross-entropy loss measurement comprises:
-
∑
n
i
,
n
j
y
n
i
,
n
j
log
(
y
ˆ
n
i
,
n
j
)
+
(
1
-
y
n
i
,
n
j
)
log
(
1
-
y
ˆ
n
i
,
n
j
)
where ŷ n i, n j represents a calculated edge value for a node pair n i , n j and ŷ n i, n j represents a predicted edge value for the node pair input for the node pair n i ,n j .
6. The non-transitory machine-readable medium of claim 1 , where the neural network comprises a scalable graph neural network.
7. A process comprising:
harvesting geo-tagged ground-level images through a scene scene detection algorithm executed by an image classifier;
receiving point of interest data representing structures or events through a transceiver;
tagging each of the geo-tagged ground-level images with a probability object describing a tag classification;
tagging each of the point of interest data with a hierarchical level genre classification object;
encoding tagged geo-tagged ground-level images as encoded vectors to render a first plurality of nodes and a first edge on a proximity graph;
encoding tagged point of interest data as a similarity vector to render a second plurality of nodes and a second edge on the proximity graph associated with the first node by a calculated semantic distance measurement;
splitting the proximity graph into a training subgraph and a testing subgraph;
initiating a training session of a neural network by processing the training subgraph during a training session that includes aggregating information at a selected node within the training subgraph by sampling information from a neighboring node within the training subgraph; and
terminating the training session when a binary cross-entropy loss measurement based on the training subgraph falls below a predetermined threshold.
8. The process of claim 7 , where the calculated semantic distance measurement comprises a calculated Euclidean semantic distance measurement.
9. The process of claim 7 , further comprising initiating an evaluation session of the neural network by processing the neural network processing the testing subgraph.
10. The process of claim 7 , where the proximity graph is constructed based on a dot product.
11. The process of claim 7 , where the binary cross-entropy loss measurement comprises an average loss measurement based on a predicted edge value and a calculated edge value.
12. The process of claim 7 , where the neural network comprises a scalable graph neural network.
13. An edge prediction system comprising:
an image classifier that harvests geo-tagged ground-level images through one or more scene scene detection algorithms;
a transceiver that receives point of interest data representing structures or events through a transceiver;
a processor that tags each of the geo-tagged ground-level images with a probability object describing a tag classification and tags each of the point of interest data with one of a plurality of hierarchical level genre classification objects;
an encoder that encodes tagged geo-tagged ground-level images as hot encoded vectors to render a first plurality of nodes and a first edge on a proximity graph, and encodes tagged point of interest data as a similarity vector to render a second plurality of nodes and a second edge on the proximity graph associated with the first plurality of nodes by a calculated semantic distance measurement;
a partitioning engine that splits the proximity graph into a training subgraph and a testing subgraph;
a training engine that processes the training subgraph through a neural network during a training session that includes aggregating information at a plurality of selected nodes within the training subgraph by sampling information from one more neighboring node within the training subgraph, and terminates the training session when a binary cross-entropy loss measurement based on the training subgraph falls below a predetermined threshold; and
a testing engine that evaluates the neural network by processing the testing subgraph.
14. The system of claim 13 , where the calculated semantic distance measurement comprises a calculated Euclidean semantic distance measurement.
15. The system of claim 13 , where the binary cross-entropy loss measurement comprises an average loss measurement based on a plurality of predicted edge values and a plurality of calculated edge values.
16. The system of claim 15 , where the binary cross-entropy loss measurement is expressed by:
-
∑
n
i
,
n
j
y
n
i
,
n
j
log
(
y
ˆ
n
i
,
n
j
)
+
(
1
-
y
n
i
,
n
j
)
log
(
1
-
y
ˆ
n
i
,
n
j
)
where y n i, n j represents a calculated edge value for a node pair n i , n j and ŷ n i, n j represents a predicted edge value for the node pair input for the node pair n i ,n j .
17. The system of claim 13 , where the proximity graph is constructed based on a dot product of the first plurality of nodes and the second plurality of nodes.