Scale Ai, Inc.

Chiao-Lun Cheng

Leigh Marie Braswell

Elliot Branson

Nathaniel John Herman

Evan Moss

Akshat Bubna

Steven Hao

Samuel Jacob Clearman

Guido Leandro Maliandi

Adrian Yunpfei Lam

Alexandre Matton

Shariq Shahab Hashme

Jeffrey Alan Anders

Anastasiia Alokhina

Suchir Balaji

Andrew Liu

Simon Alexander Hewat

Nathan Hayflick

Nathan Herman

Daniel Havír

Daniel Havíř

One embodiment of the present invention sets forth a technique for extracting data from a document. The technique includes determining a first set of features associated with the document, wherein the first set of features comprises a set of region proposals that bound one or more portions of text within the document. The technique also includes applying a first machine learning model to the first set of features to generate a set of predictions associated with one or more key-value pairs included in the document. The technique further includes extracting the one or more key-value pairs from the document based on the set of predictions.

Key-value extraction from documents

One embodiment of the present invention sets forth a technique for extracting data from a document. The technique includes determining, via execution of one or more machine learning models, a first set of bounding boxes for a first set of values associated with a first key within the document and a second set of bounding boxes for a second set of values associated with a second key within the document. The technique also includes generating a first set of mappings between a list of items in the document and the first set of bounding boxes and a second set of mappings between the first and second sets of bounding boxes based on locations of the bounding boxes. The technique further includes determining, for a given item, one or more associated bounding boxes in the first and second sets of bounding boxes based on the mappings.

Linking key-value pairs in documents

One embodiment of the present invention sets forth a technique for performing a labeling task. The technique includes determining one or more region proposals, wherein each region proposal included in the one or more region proposals includes estimates of one or more bounding boxes surrounding one or more objects in a plurality of video frames. The technique also includes performing one or more operations that execute a refinement stage of a machine learning model to produce one or more refined estimates of the one or more bounding boxes included in the one or more region proposals. The technique further includes outputting the one or more refined estimates as initial representations of the one more bounding boxes for subsequent annotation of the one or more bounding boxes by one or more users.

Prelabeling of bounding boxes in video frames

One embodiment provides techniques for automatically pre-labeling point cloud data with cuboid annotations. Point cloud data is processed using ML models to detect, associate, and localize objects therein, in order to generate cuboid tracks that each include a series of cuboid annotations associated with an object. An object detection model that detects objects and performs coarse localization is trained using a loss function that separately evaluates the distances between corners of predicted cuboids and corners of ground truth cuboids for position, size, and yaw. A refinement model that performs more accurate localization takes as input 2D projections of regions surrounding cuboid tracks predicted by the object detection model and the cuboid tracks, and outputs refined cuboid tracks. The refined cuboid tracks are filtered to a set of keyframes, with in-between frames being interpolated. The cuboid tracks can then be presented to a user for viewing and editing.

Pre-labeling data with cuboid annotations

One embodiment of the present invention sets forth a technique for performing a labeling task. The technique includes generating a multi-scale representation of an image as input to a machine learning model. The technique also includes performing one or more operations that apply the machine learning model to the multi-scale representation of the image to produce a semantic segmentation comprising predictions of labels for regions of pixels in the image. The technique further includes outputting, in a user interface, the semantic segmentation for use in assisting a user in specifying the labels for the pixels in the image.

Prelabeling for semantic segmentation tasks

One embodiment provides a user interface (UI) that permits users to select how point cloud colorings determined from multiple data sources are blended together in a rendering of a point cloud. The data sources may include photographic, label, and/or LIDAR intensity data. To improve frame rates, an aggregated point cloud may be generated using a spatial hash of a large set of points and sampling of each hash bucket based on the number of points therein and a user-configurable density. Sizes of points in the point cloud may decrease proportionally to distance from a viewer, but increase based on an activation function that enlarges points greater than a threshold distance from the viewer. In addition, luminance statistics for sub-regions of photographic data and dominant colors determined from photographic data may be used to automatically determine color properties to apply to a point cloud coloring.

Visualization techniques for data labeling

Growing community of inventors

San Francisco, CA, United States of America

Chiao-Lun Cheng