Map-Free Visual Relocalization

1 . A computer-implemented method for providing map-free relocalization of a device, the method comprising:

obtaining a reference image of an environment captured by a reference camera at a reference pose;

receiving a query image taken by a camera of the device;

applying the reference image and the query image to a relative pose regression network to output a relative pose of the camera of the device relative to the reference camera in the environment, the relative pose regression network comprising:

a Siamese network configured to receive the reference image to generate a first set of feature maps, and receive the query image to generate a second set of feature maps;

a correlation network configured to receive the first set of feature maps and the second set of feature maps as input to generate a set of global features;

a residual network configured to receive the set of global features as input to generate a global feature vector; and

a multilayer perceptron network configured to receive the global feature vector as input to determine the relative pose of the camera of the device relative to the reference camera in the environment; and

determining a pose of the camera of the device based on the relative pose of the camera of the device and the reference pose of the reference camera.

2 . The method of claim 1 , wherein the relative pose is a first relative pose indicating a pose of the camera of the device relative to the reference camera, the method comprising:

applying the query image and the reference image to the relative pose regression network to generate a second relative pose indicating a pose of the reference camera relative to the camera of the device;

determining a third relative pose indicating a pose of the camera of the device relative to the reference camera based on the second relative pose; and

determining an updated relative pose based on the first relative pose and the third relative pose.

3 . The method of claim 1 , wherein Siamese network comprises a first deep residual UNET and a second deep residual UNET, each of which is configured to receive the reference image or the query image:

4 . The method of claim 1 , wherein the relative pose regression network further comprising:

a second multilayer perceptron network configured to receive the global feature vector as input to generate an angular error indicating a confidence level of the determined relative pose.

5 . The method of claim 4 , wherein the angular error is determined with respect to a ground truth relative pose.

6 . The method of claim 4 , wherein second multiplayer perceptron network is trained based on a soft clamping function, describing a ground truth error and a network prediction.

7 . The method of claim 1 , wherein the correlation network is configured to compute a 4-dimensional correlation volume to mimic soft feature matching.

8 . The method of claim 7 , wherein the correlation network is further configured to use the 4-dimensional correlation volume to warp the second set of feature maps and a regular grid of coordinates.

9 . The method of claim 8 , wherein the correlation network is further configured to concatenate the warped second set of feature maps and the warped regular grid of coordinates to generate the set of global features.

10 . The method of claim 8 , wherein the relative pose regression network parameterizes rotations as a plurality of discrete angles.

11 . The method of claim 8 , wherein the relative pose regression network is trained via a training dataset comprising a plurality of pairs of training images, each pair of training images are taken in a same environment.

12 . The method of claim 11 , wherein each training image is associated with an absolute pose.

13 . The method of claim 11 , wherein each pair of training images is associated with an overlap score, indicating a level of overlapping between the pair of training images.

14 . The method of claim 11 , wherein each training image is associated with camera intrinsics describing information associated with a camera that took the training image.

15 . The method of claim 11 , wherein each training image is anonymized by detecting and blurring personally identifiable information on the training image.

16 . A computer-implemented method for providing map-free relocalization of a device, the method comprising:

obtaining a reference image of an environment captured by a reference camera, wherein the reference image associated with a pose of the reference camera in the environment;

capturing a query image of the environment by a camera of the device;

generating a first depth map of the reference image;

generating a second depth map of the query image;

determining depth correspondences between the first depth map and the second depth map;

determining a pose of the camera of the device based in part on the depth correspondences between the first depth map and the second depth map.

17 . The method of claim 16 , the method further comprising:

determining 2-dimensional to 2-dimensional (2D-2D) correspondences between 2-dimensional points on the query image and 2-dimensional points on the reference image; and

determining the pose of the camera of the device further based on 2D-2D correspondences.

18 . The method of claim 16 , the method further comprising:

back-projecting one of first depth map or second depth map to a 3-dimensional image;

determining 2-dimensional to 3-dimensional (2D-3D) correspondences between the query image or the reference image and the 3-dimensional image;

determining a pose of the camera of the device further based on the 2D-3D correspondences.

19 . The method of claim 16 , the method further comprising:

back-projecting the reference image to a first 3-dimensional image based on the first depth map;

back-projecting the query image to a second 3-dimensional image based on the second depth map;

determining 3-dimensional to 3-dimensional (3D-3D) correspondences between 3-dimensional points on the first 3-dimensional image and 3-dimensional points on the second 3-dimensional image, wherein each 3D-3D correspondence provides a scale estimate for a translation vector; and

determining a pose of the camera of the device further based on the 3D-3D correspondences.

20 . The method of claim 16 , further comprising:

accessing a dataset comprising a plurality of reference images to obtain the reference image, each of the plurality of reference images is an image of a place of interest that is well captured by a single image.