Honda Motor Co., Ltd.

Siemens Corporation

Siemens Corporate Research, Inc.

Larry Davis

Teruhisa Misu

Yi-Ting Chen

Anurag Mittal

Ahmed Taha

Ser-Nam Lim

Xitong Yang

Multi-modal data representing driving events and corresponding actions related to the driving events can be obtained and used to train a neural network at least in part by using a triplet loss computed for the driving events as a regression loss to determine an embedding of driving event data. In some cases, using the trained neural network, a retrieval request for an input driving event and corresponding action can be processed by determining, from the neural network, one or more similar driving events or corresponding actions in the multi-modal data.

Systems for modeling uncertainty in multi-modal retrieval and methods thereof

Methods, systems, and computer-readable mediums storing computer executable code for visual recognition implementing a triplet loss function are provided. The method include receiving an image generated from an image source associated with a vehicle. The method may also include analyzing the image based on a convolutional neural network. The convolutional neural network may apply both a triplet loss function and a softmax loss function to the image to determine classification logits. The method may also include classifying the image into a predetermined class distribution based upon the determined classification logits. The method may also include instructing the vehicle to perform a specific task based upon the classified image.

Methods and systems for visual recognition using triplet loss

A multi-camera system that collects images and videos of moving objects in dynamic and crowded scenes, subject to task constraints is disclosed. The system constructs 'task visibility intervals' comprising information about what can be sensed in future time intervals. Methods for constructing these intervals applying prediction of future object motion and including consideration of factors such as object occlusion and camera control parameters are also disclosed. Using a plane-sweep algorithm, these atomic intervals can be combined in a method to form multi-task intervals, during which a single camera can collect videos suitable for multiple tasks simultaneously. Methods for fast camera scheduling that yield solutions within a small constant factor of an optimal solution are also disclosed.

Smart task-driven video collection

A method for eliminating errors in foreground object detection in digitized images comprises providing a reference camera and a secondary camera, vertically aligning each said camera with a baseline that is approximately perpendicular to a ground plane, wherein said reference camera is placed lower than said secondary camera, selecting a foreground pixel in a reference view of a first point in a foreground object, finding a conjugate pixel of the foreground pixel in a secondary view, using the foreground and conjugate pixels to determine an image base pixel of a base point in the reference view, wherein said base point is a point on the ground plane below the first point, and using the foreground and image base pixels to find a location where the ground plane is first visible.

Growing community of inventors

Potomac, MD, United States of America

Larry Davis