Five AI Limited

The present disclosure relates to a computer-implemented method of training a generative model to insert an object in spatial sensor data. The method comprises receiving a training sample of spatial sensor data and receiving an indication of a 3D geometric property of an object captured in the training sample. A portion of spatial sensor data corresponding to the object is removed from the training sample, resulting a cropped training sample. The generative model is trained to reconstruct the training sample from the cropped training sample by, providing to the generative model: the cropped training sample as a target input, an indication of the object as a reference input, and the 3D geometric property of the object as a conditioning input. This results in a generated output sample of spatial sensor data. Parameters of the generative model are tuned to reduce a reconstruction error between the training sample and the generated output sample, resulting in a trained generative model configured to insert at inference, in a set of spatial sensor data received as a target input, an object indicated by a reference input with a desired 3D geometric property indicated by a conditioning input.

The present disclosure relates to techniques for training a generative model to insert an object in spatial sensor data. A first training sample of spatial sensor data of a first sensor modality, and a second training sample of spatial sensor data of a second sensor modality are received, the first training sample and the second training sample capture a common object. A first portion of sensor data corresponding to the object is removed from the first training sample, resulting a cropped training sample. A second portion of spatial sensor data corresponding to the common object is extracted from the second training sample. The generative model is trained to reconstruct the first training sample from the cropped training sample by: providing to the generative model: the cropped training sample as a target input, and the second portion of spatial sensor data as a reference input, resulting in a generated output sample of spatial sensor data, and tuning parameters of the generative model to reduce a reconstruction error between the first training sample and the generated output sample. This results in a trained generative model configured to insert at inference, in a first set of spatial sensor data of the first modality received as a target input, an object indicated in a second set of spatial sensor data of the second modality received as a reference input.

To locate and model a 3D object captured in multiple time-series of sensor data of multiple sensor modalities, a cost function applied to the multiple time-series of sensor data is optimized. The cost function aggregates over time and the multiple sensor modalities, and is defined over a set of variables comprising one or more shape parameters of a 3D object model and a time sequence of poses of the 3D object model. The cost function penalizes inconsistency between the multiple time-series of sensor data and the set of variables. The object belongs to a known object class, and the 3D object model or the cost function encodes expected 3D shape information associated with the known object class, whereby the 3D object is located at multiple time instants and modelled by tuning each pose and the shape parameters with the objective of optimizing the cost function.

An encoder is trained together with a perception component based on a training set comprising unannotated sensor data sets and annotated sensor data sets in a sequence of multiple training steps. Each training step comprises: in a first phase of the training step, updating the set of encoder parameters based on the unannotated sensor data sets, with the aim of optimizing a self-supervised loss function, without updating the set of task-specific parameters of the perception component, and in a second phase of the training step, updating the set of task-specific parameters based on the annotated sensor data sets, with the aim of optimizing a task-specific loss function, wherein the encoder as updated in the first phase of that training step processes a data representation of each annotated sensor data set to extract features therefrom, wherein the perception component processes the extracted features to compute an output therefrom, and wherein the task-specific loss is defined on the output and the associated annotation for each annotated sensor data set for learning a desired perception task. In performing the sequence of multiple training steps, the method alternates repeatedly between the first phase and the second phase.

A computer-implemented method of perceiving structure in a radar point cloud comprises: generating a discretised image representation of the radar point cloud having (i) an occupancy channel indicating whether or not each pixel of the discretised image representation corresponds to a point in the radar point cloud and (ii) a Doppler channel containing, for each occupied pixel, a Doppler velocity of the corresponding point in the radar point cloud; and inputting the discretised image representation to a machine learning (ML) perception component, which has been trained extract information about structure exhibited in the radar point cloud from the occupancy and Doppler channels.

In one aspect, a method of extracting depth information from a stereoscopic pair of images comprises: in a first depth extraction phase, determining a set of support points in 3D space for at least a first image of the stereoscopic image pair; using the set of support points to determine a 3D surface prior; and in a second depth extraction phase, computing, for the first image, a second-stage disparity map, by optimizing a second cost function defined so as to penalize deviation in the second-stage disparity map from the 3D surface prior. Disparity smoothness may be encouraged in at least one of the first and second depth extraction phases using a cost function that penalizes deviations in disparity between nearby pixels.

A computer implemented method of training an encoder to extract features from sensor data comprises generating a plurality of training examples, each training example comprising at least two data representations of a set of sensor data, the at least two data representations related by a transformation parameterized by at least one numerical transformation value; and training the encoder based on a self-supervised regression loss function applied to the training examples. The encoder extracts respective features from the at least two data representations of each training example, and at least one numerical output value is computed from the extracted features. The self-supervised regression loss function encourages the at least one numerical output value to match the at least one numerical transformation value parameterizing the transformation.

An occlusion metric is computed for a target object in a 3D multi-object simulation. The target object is represented in 3D space by a collision surface and a 3D bounding box. In a reference surface defined in 3D space, a bounding box projection is determined for the target object with respect to an ego location. The bounding box projection is used to determine a set of reference points in 3D space. For each reference point of the set of reference points, a corresponding ray is cast based on the ego location, and it is determined whether the ray is an object ray that intersects the collision surface of the target object. For each such object ray, it is determined whether the object ray is occluded. The occlusion metric conveys an extent to which the object rays are occluded.

A computer-implemented method of modelling a common structure component, the method comprising, in a modelling computer system: receiving a plurality of captured frames, each frame comprising a set of 3D structure points, in which at least a portion of a common structure component is captured; computing a first reference position within at least one first frame of the plurality of frames; selectively extracting first 3D structure points of the first frame based on the first reference position computed for the first frame; computing a second reference position within a second frame of the plurality of frames; selectively extracting second 3D structure points of the second frame based on the second reference position computed for the second frame; and aggregating the first 3D structure points and the second 3D structure points, thereby generating an aggregate 3D model of the common structure component based on the first and second reference positions.

A computer-implemented method of training a depth uncertainty estimator comprises receiving, at a training computer system, a set of training examples, each training example comprising (i) a stereo image pair and (ii) an estimated disparity map computed from at least one image of the stereo image pair by a depth estimator. The training computer system executes a training process to learn one or more uncertainty estimation parameters of a perturbation function, the uncertainty estimation parameters for estimating uncertainty in disparity maps computed by the depth estimator. The training process is performed by sampling a likelihood function based on the training examples and the perturbation function, thereby obtaining a set of sampled values for learning the one or more uncertainty estimation parameters. The likelihood function measures similarity between one image of each training example and a reconstructed image computed by transforming the other image of that training example based on a possible true disparity map derived from the estimated disparity map of that training example and the perturbation function.