CROSS REALITY SYSTEM FOR LARGE SCALE ENVIRONMENTS
A cross reality system enables any of multiple devices to efficiently and accurately access previously persisted maps of very large scale environments and render virtual content specified in relation to those maps. The cross reality system may quickly determine whether a 2D set of features derived from images acquired with a portable device match a set of 3D features of an environment map and, if so, determine the relative pose of the feature sets. The pose may be used in quickly and accurately localizing the portable device to the environment map. Pairs of features in the 2D and 3D features sets may be identified based on matching feature descriptors and may be scored in a neural network trained to assess the quality of the match. Poses may be identified based on subsets of the matching features weighted towards pairs of features with high quality.
1 . (canceled)
2 . A method of aligning a first set of features, derived from at least one image collected on a portable electronic device, at least partially to a second set of features in a stored map, the method comprising:
identifying a plurality of pairs of matched features from the first and second sets of features, each of the plurality of pairs of matched features comprising a first feature of the first set and a second feature of the second set;
for each of the plurality of pairs of matched features, computing a quality metric; and
using the plurality of pairs of matched features and respective quality metrics for aligning the first set of features to the second set of features by randomly selecting and processing subsets of matched feature pairs until a stop condition is reached,
wherein:
the quality metric for a pair of matched features indicates a likelihood of finding an alignment between the first set of features and the second set of features with a subset of the plurality of pairs of matched features if the pair of matched features are included in the subset of the plurality of pairs of matched features; and
randomly selecting comprises a random selection weighted based on the quality metrics of the plurality of pairs.
3 . The method of claim 2 , wherein the computing is by a trained neural network configured for:
computing, based on feature information indicating a position of the first feature in a first coordinate frame of the portable electronic device and a position of the second feature in a second coordinate frame of the stored map, encoded vectors having a dimensionality greater than the feature information;
computing, based on the encoded vectors, decoded vectors having a dimensionality less than the encoded vectors; and
computing, based on the decoded vectors, the quality metrics for respective pairs of the plurality of pairs of matched features.
4 . The method of claim 3 , wherein the feature information comprises:
a first vector indicating a position of the first feature in a sensor coordinate frame of a sensor captured an image containing the first feature;
a second vector indicating a position of the sensor in the first coordinate frame; and
a third vector indicating a position of the second feature in the second coordinate frame.
5 . The method of claim 4 , wherein:
the trained neural network comprises a first embedding layer, a second embedding layer, and an intermediate layer between the first and second embedding layers; and
the first embedding layer comprises a plurality of first multi-layer perceptrons each configured for computing an encoded vector based on the feature information of one or more pairs of the plurality of pairs of matched features.
6 . The method of claim 5 , wherein:
the second embedding layer comprises a plurality of second multi-layer perceptrons each configured for computing a decoded vector based on a respective encoded vector.
7 . The method of claim 6 , wherein:
the trained neural network comprises a plurality of activation blocks configured for converting the decoded vectors into the quality metrics.
8 . The method of claim 3 , wherein:
the identifying is based on descriptors of the features of the first and second sets;
the trained neural network is a first trained neural network; and
the method comprises computing, with a second trained neural network, descriptors for the features of the first set.
9 . An electronic system that supports specification of a position of virtual content relative to stored maps in a database of stored maps, the system comprising:
a communication component configured to receive from a portable electronic device information about a first set of features in a three-dimensional (3D) environment of the portable electronic device; and
a computing component, connected to the communication component, the computing component configured to:
identify a plurality of pairs of matched features from the first set of features and a second set of features in a stored map of the database of stored maps, each of the plurality of pairs of matched features comprising a first feature of the first set and a second feature of the second set;
for each of the plurality of pairs of matched features, compute a quality metric based on feature information indicating a position of the first feature in a first coordinate frame of the portable electronic device and a position of the second feature in a second coordinate frame of the stored map; and
provide the plurality of pairs of matched features and respective quality metrics for aligning the first set of features at least partially to the second set of features.
10 . The system of claim 9 , wherein the computing component comprises a trained neural network configured for:
computing, based on the feature information, encoded vectors having a dimensionality greater than the feature information;
computing, based on the encoded vectors, decoded vectors having a dimensionality less than the encoded vectors; and
computing, based on the decoded vectors, the quality metrics for respective pairs of the plurality of pairs of matched features.
11 . The system of claim 10 , wherein the feature information comprises:
a first vector indicating a position of the first feature in a sensor coordinate frame of a sensor captured an image containing the first feature;
a second vector indicating a position of the sensor in the first coordinate frame; and
a third vector indicating a position of the second feature in the second coordinate frame.
12 . The system of claim 11 , wherein:
the trained neural network comprises a first embedding layer, a second embedding layer, and an intermediate layer between the first and second embedding layers; and
the first embedding layer comprises a plurality of first multi-layer perceptrons each configured for computing an encoded vector based on the feature information of one or more pairs of the plurality of pairs of matched features.
13 . The system of claim 12 , wherein:
the second embedding layer comprises a plurality of second multi-layer perceptrons each configured for computing a decoded vector based on a respective encoded vector.
14 . The system of claim 13 , wherein:
the trained neural network comprises a plurality of activation blocks configured for converting the decoded vectors into the quality metrics.
15 . The system of claim 10 , wherein:
the trained neural network is a first trained neural network;
the computing component comprises a second trained neural network configured to compute descriptors for the features of the first set; and
the plurality of pairs of matched featured are identified based on the descriptors of the features of the first and second sets.
16 . A non-transitory computer-readable medium storing computer executable instructions configured to, when executed by at least one processor, perform a method of aligning a first set of features, derived from at least one image collected on a portable electronic device, at least partially to a second set of features in a stored map, the method comprising:
identifying a plurality of pairs of matched features from the first and second sets of features, each of the plurality of pairs of matched features comprising a first feature of the first set and a second feature of the second set;
for each of the plurality of pairs of matched features, computing a quality metric based on feature information indicating a position of the first feature in a first coordinate frame of the portable electronic device and a position of the second feature in a second coordinate frame of the stored map; and
providing the plurality of pairs of matched features and respective quality metrics for aligning the first set of features at least partially to the second set of features.
17 . The non-transitory computer-readable medium of claim 16 , wherein the computer executable instructions comprise a trained neural network configured for:
computing, based on the feature information, encoded vectors having a dimensionality greater than the feature information;
computing, based on the encoded vectors, decoded vectors having a dimensionality less than the encoded vectors; and
computing, based on the decoded vectors, the quality metrics for respective pairs of the plurality of pairs of matched features.
18 . The non-transitory computer-readable medium of claim 17 , wherein the feature information comprises:
a first vector indicating a position of the first feature in a sensor coordinate frame of a sensor captured an image containing the first feature;
a second vector indicating a position of the sensor in the first coordinate frame; and
a third vector indicating a position of the second feature in the second coordinate frame.
19 . The non-transitory computer-readable medium of claim 18 , wherein:
the trained neural network comprises a first embedding layer, a second embedding layer, and an intermediate layer between the first and second embedding layers; and
the first embedding layer comprises a plurality of first multi-layer perceptrons each configured for computing an encoded vector based on the feature information of one or more pairs of the plurality of pairs of matched features.
20 . The non-transitory computer-readable medium of claim 19 , wherein:
the second embedding layer comprises a plurality of second multi-layer perceptrons each configured for computing a decoded vector based on a respective encoded vector.
21 . The non-transitory computer-readable medium of claim 20 , wherein:
the trained neural network comprises a plurality of activation blocks configured for converting the decoded vectors into the quality metrics.