Abstract: An image matching system for determining visual overlaps between images by using box embeddings is described herein. The system receives two images depicting a 3D surface with different camera poses. The system inputs the images (or a crop of each image) into a machine learning model that outputs a box encoding for the first image and a box encoding for the second image. A box encoding includes parameters defining a box in an embedding space. Then the system determines an asymmetric overlap factor that measures asymmetric surface overlaps between the first image and the second image based on the box encodings. The asymmetric overlap factor includes an enclosure factor indicating how much surface from the first image is visible in the second image and a concentration factor indicating how much surface from the second image is visible in the first image.

Abstract: An image matching system for determining visual overlaps between images by using box embeddings is described herein. The system receives two images depicting a 3D surface with different camera poses. The system inputs the images (or a crop of each image) into a machine learning model that outputs a box encoding for the first image and a box encoding for the second image. A box encoding includes parameters defining a box in an embedding space. Then the system determines an asymmetric overlap factor that measures asymmetric surface overlaps between the first image and the second image based on the box encodings. The asymmetric overlap factor includes an enclosure factor indicating how much surface from the first image is visible in the second image and a concentration factor indicating how much surface from the second image is visible in the first image.

Abstract: A model predicts the geometry of both visible and occluded traversable surfaces from input images. The model may be trained from stereo video sequences, using camera poses, per-frame depth, and semantic segmentation to form training data, which is used to supervise an image to image network. In various embodiments, the model is applied to a single RGB image depicting a scene to produce information describing traversable space of the scene that includes occluded traversable. The information describing traversable space can include a segmentation mask of traversable space (both visible and occluded) and non-traversable space and a depth map indicating an estimated depth to traversable surfaces corresponding to each pixel determined to correspond to traversable space.

Abstract: A computer system generates stereo image data from monocular images. The system generates depth maps for single images using a monocular depth estimation method. The system converts the depth maps to disparity maps and uses the disparity maps to generate additional images forming stereo pairs with the monocular images. The stereo pairs can be used to form a stereo image training data set for training various models, including depth estimation models or stereo matching models.

Abstract: The invention relates to a control device (1) for a vehicle for determining the perceptual load of a visual and dynamic driving scene. The control device is configured to: ?receive a sensor output (101) of a sensor (3), the sensor (3) sensing the visual driving scene, ?extract a set of scene features (102) from the sensor output (101), the set of scene features (102) representing static and/or dynamic information of the visual driving scene, ?determine the perceptual load (104) of the set of extracted scene features (102) based on a predetermined load model (103), the load model (103) being predetermined based on reference video scenes each being labelled with a load value, ?map the perceptual load to the sensed driving and ?determine a spatial and temporal intensity distribution of the perceptual load across the sensed driving scene. The invention further relates to a vehicle, a system and a method.

Abstract: A parallel-reality game uses a virtual game board having tiles placed over an identified traversable space corresponding to flat regions of a scene. A game board generation module receives one or more images of the scene captured by a camera of a mobile device. The game board generation module obtains a topographical mesh of the scene based on the received one or more images. The game board generation module then identifies a traversable space within the scene based on the obtained topographical mesh. The game board generation module determines a location for each of a set of polygon tiles in the identified traversable space. The game board generation module also allows for queries to identify parts of the game board that meet one or more provided criterion.

Abstract: A model predicts the geometry of both visible and occluded traversable surfaces from input images. The model may be trained from stereo video sequences, using camera poses, per-frame depth, and semantic segmentation to form training data, which is used to supervise an image to image network. In various embodiments, the model is applied to a single RGB image depicting a scene to produce information describing traversable space of the scene that includes occluded traversable. The information describing traversable space can include a segmentation mask of traversable space (both visible and occluded) and non-traversable space and a depth map indicating an estimated depth to traversable surfaces corresponding to each pixel determined to correspond to traversable space.

Abstract: A method for training a depth estimation model with depth hints is disclosed. For each image pair: for a first image, a depth prediction is determined by the depth estimation model and a depth hint is obtained; the second image is projected onto the first image once to generate a synthetic frame based on the depth prediction and again to generate a hinted synthetic frame based on the depth hint; a primary loss is calculated with the synthetic frame; a hinted loss is calculated with the hinted synthetic frame; and an overall loss is calculated for the image pair based on a per-pixel determination of whether the primary loss or the hinted loss is smaller, wherein if the hinted loss is smaller than the primary loss, then the overall loss includes the primary loss and a supervised depth loss between depth prediction and depth hint. The depth estimation model is trained by minimizing the overall losses for the image pairs.

Abstract: A scene reconstruction model is disclosed that outputs a heightfield for a series of input images. The model, for each input image, predicts a depth map and extracts a feature map. The model builds a 3D model utilizing the predicted depth maps and camera poses for the images. The model raycasts the 3D model to determine a raw heightfield for the scene. The model utilizes the raw heightfield to sample features from the feature maps corresponding to positions on the heightfield. The model aggregates the sampled features into an aggregate feature map. The model regresses a refined heightfield based on the aggregate feature map. The model determines the final heightfield based on a combination of the raw heightfield and the refined heightfield. With the final heightfield, a client device may generate virtual content augmented on real-world images captured by the client device.

Abstract: Panoptic segmentation forecasting predicts future positions of foreground objects and background objects separately. An egomotion model may be implemented to estimate egomotion of the camera. Pixels in frames of captured video are classified between foreground and background. The foreground pixels are grouped into foreground objects. A foreground motion model forecasts motion of the foreground objects to a future timestamp. A background motion model backprojects the background pixels into point clouds in a three-dimensional space. The background motion model predicts future positions of the point clouds based on egomotion. The background motion model may further generate novel point clouds to fill in occluded space. With the predicted future positions, the foreground objects and the background pixels are combined into a single panoptic segmentation forecast.

Abstract: A method for training a depth estimation model with depth hints is disclosed. For each image pair: for a first image, a depth prediction is determined by the depth estimation model and a depth hint is obtained; the second image is projected onto the first image once to generate a synthetic frame based on the depth prediction and again to generate a hinted synthetic frame based on the depth hint; a primary loss is calculated with the synthetic frame; a hinted loss is calculated with the hinted synthetic frame; and an overall loss is calculated for the image pair based on a per-pixel determination of whether the primary loss or the hinted loss is smaller, wherein if the hinted loss is smaller than the primary loss, then the overall loss includes the primary loss and a supervised depth loss between depth prediction and depth hint. The depth estimation model is trained by minimizing the overall losses for the image pairs.

Abstract: A multi-frame depth estimation model is disclosed. The model is trained and configured to receive an input image and an additional image. The model outputs a depth map for the input image based on the input image and the additional image. The model may extract a feature map for the input image and an additional feature map for the additional image. For each of a plurality of depth planes, the model warps the feature map to the depth plane based on relative pose between the input image and the additional image, the depth plane, and camera intrinsics. The model builds a cost volume from the warped feature maps for the plurality of depth planes. A decoder of the model inputs the cost volume and the input image to output the depth map.

Abstract: The present disclosure describes approaches to camera re-localization using a graph neural network (GNN). A re-localization model includes encoding an input image into a feature map. The model retrieves reference images from an image database of a previously scanned environment based on the feature map of the image. The model builds a graph based on the image and the reference images, wherein nodes represent the image and the reference images, and edges are defined between the nodes. The model may iteratively refine the graph through auto-aggressive edge-updating and message passing between nodes. With the graph built, the model predicts a pose of the image based on the edges of the graph. The pose may be a relative pose in relation to the reference images, or an absolute pose.

Abstract: A method for training a depth estimation model with depth hints is disclosed. For each image pair: for a first image, a depth prediction is determined by the depth estimation model and a depth hint is obtained; the second image is projected onto the first image once to generate a synthetic frame based on the depth prediction and again to generate a hinted synthetic frame based on the depth hint; a primary loss is calculated with the synthetic frame; a hinted loss is calculated with the hinted synthetic frame; and an overall loss is calculated for the image pair based on a per-pixel determination of whether the primary loss or the hinted loss is smaller, wherein if the hinted loss is smaller than the primary loss, then the overall loss includes the primary loss and a supervised depth loss between depth prediction and depth hint. The depth estimation model is trained by minimizing the overall losses for the image pairs.

Abstract: An image localization system receives an image of a scene and generates a depth map for the image by inputting the image to a model trained for generating depth maps for images. The system determines surface normal vectors for the pixels in the depth map. The system clusters the surface normal vectors to identify regions in the image corresponding to planar surfaces. The system partitions the image into patches, each of which is a region of connected pixels in the image and corresponds to a cluster of surface normal vectors. The system rectifies the perspective distortion of patches and extracts perspective corrected features from the rectified patches. The system matches the perspective corrected features of the image with perspective corrected features of other images for three-dimensional re-localization.

Abstract: An image matching system for determining visual overlaps between images by using box embeddings is described herein. The system receives two images depicting a 3D surface with different camera poses. The system inputs the images (or a crop of each image) into a machine learning model that outputs a box encoding for the first image and a box encoding for the second image. A box encoding includes parameters defining a box in an embedding space. Then the system determines an asymmetric overlap factor that measures asymmetric surface overlaps between the first image and the second image based on the box encodings. The asymmetric overlap factor includes an enclosure factor indicating how much surface from the first image is visible in the second image and a concentration factor indicating how much surface from the second image is visible in the first image.

Abstract: A computer system generates stereo image data from monocular images. The system generates depth maps for single images using a monocular depth estimation method. The system converts the depth maps to disparity maps and uses the disparity maps to generate additional images forming stereo pairs with the monocular images. The stereo pairs can be used to form a stereo image training data set for training various models, including depth estimation models or stereo matching models.

Abstract: A method for training a depth estimation model and methods for use thereof are described. Images are acquired and input into a depth model to extract a depth map for each of the plurality of images based on parameters of the depth model. The method includes inputting the images into a pose decoder to extract a pose for each image. The method includes generating a plurality of synthetic frames based on the depth map and the pose for each image. The method includes calculating a loss value with an input scale occlusion and motion aware loss function based on a comparison of the synthetic frames and the images. The method includes adjusting the plurality of parameters of the depth model based on the loss value. The trained model can receive an image of a scene and generate a depth map of the scene according to the image.

Abstract: A model predicts the geometry of both visible and occluded traversable surfaces from input images. The model may be trained from stereo video sequences, using camera poses, per-frame depth, and semantic segmentation to form training data, which is used to supervise an image to image network. In various embodiments, the model is applied to a single RGB image depicting a scene to produce information describing traversable space of the scene that includes occluded traversable. The information describing traversable space can include a segmentation mask of traversable space (both visible and occluded) and non-traversable space and a depth map indicating an estimated depth to traversable surfaces corresponding to each pixel determined to correspond to traversable space.

Abstract: A method for training a depth estimation model and methods for use thereof are described. Images are acquired and input into a depth model to extract a depth map for each of the plurality of images based on parameters of the depth model. The method includes inputting the images into a pose decoder to extract a pose for each image. The method includes generating a plurality of synthetic frames based on the depth map and the pose for each image. The method includes calculating a loss value with an input scale occlusion and motion aware loss function based on a comparison of the synthetic frames and the images. The method includes adjusting the plurality of parameters of the depth model based on the loss value. The trained model can receive an image of a scene and generate a depth map of the scene according to the image.