Samsung Electronics Co., Ltd.

Massachusetts Institute of Technology

Allan Jepson

Konstantinos Derpanis

Iqbal Ismail Mohomed

Afsaneh Fazly

Aleksai Levinshtein

Mikita Andreevich Dvornik

Haotian Zhang

Stavros Tsogkas

Isma Hadji

Ran Zhang

Caleb Ryan Phillips

Mete Kemertas

Sven Josef Dickinson

Hoda Rezaee Kaviani

Tristan Ty Aumentado-armstrong

Xinyu Sun

David J Heeger

Haicheng Wang

Vineeth Subrahmanya Bhaskara

He Zhao

Richard Wildes

A first graph which is a flow graph is converted into a second graph using a first algorithm. The second graph, combined with a second algorithm using a dynamic programming recursion, is compared with a video and a best matching thread, with respect to the video, is found through the second graph. The best matching thread may then be used for a user-assistance function. The dynamic programming recursion reduces computational effort needed by the computer performing the matching.

Grounding flow graphs in signals

The present disclosure provides methods, apparatuses, and computer-readable mediums for representing shapes with probabilistic directed distance fields. In some embodiments, a method includes obtaining a camera representation and a latent shape vector representation of a scene. The camera representation indicates position information and direction information of a view of the scene. The method further includes calculating, based on the latent shape vector representation of the scene, a visibility score and a depth for each ray of a plurality of rays emanating from a corresponding plurality of positions and directions. The plurality of positions and directions are determined from the camera representation of the scene. The method further includes generating renders of geometric information of the scene using the visibility score and the depth of the plurality of rays.

Representing 3D shapes with probabilistic directed distance fields

A method for training a super-resolution network may include obtaining a low resolution image; generating, using a first machine learning model, a first high resolution image based on the low resolution image; generating, using a second machine learning model, a second high resolution image based on the first high resolution image and an unpaired dataset of high resolution images; obtaining a training data set using the low resolution image and the second high resolution image; and training the super-resolution network using the training data set.

Unsupervised super-resolution training data construction

A method is provided. The method includes receiving a first signal and a second signal, generating a first sequence based on the first signal by embedding at least one feature in the first signal, generating a second sequence based on the second signal by embedding at least one feature in the second signal, determining a minimum cost path of aligning the first sequence with the second sequence based on the embedded at least one feature in the first sequence and the embedded at least one feature in the second sequence, and aligning the first sequence with second sequence based on the determined minimum cost path.

Method and system for learning to temporal align signals with interspersed outliers

Image alignment is achieved by obtaining a fused optical flow based on a global motion estimate and an optical flow estimate. The fused optical flow may be used to warp the first image and/or the second image into aligned features or aligned images. An output image may then be obtained by combining.

Semi-global neural image alignment

A method of personalized image retrieval includes obtaining a natural language query including a name; replacing the name in the natural language query with a generic term to provide an anonymized query and named entity information; obtaining a plurality of initial ranking scores and a plurality of attention weights corresponding to a plurality of images using a trained scoring model that inputs the anonymized query and the plurality of images; obtaining a plurality of delta scores corresponding to the plurality of images using a re-scoring model that inputs the plurality of attention weights and the named entity information; and obtaining a plurality of final ranking scores by modifying the plurality of initial ranking scores based on the plurality of delta scores. The trained scoring model performs semantic based searching and the re-scoring model determines a probability that faces detected in the plurality of images correspond to the name.

Method of personalized image and video searching based on a natural language query, and an apparatus for the same

The present disclosure provides methods and apparatuses for probabilistic procedure planning for generating a plan based on a goal relating to an end state. In some embodiments, a method includes receiving a request from a user to generate an action plan comprising T intermediate actions between a start state and the end state. The method further includes constructing an input query matrix based on T, the start state, the end state, positional encodings, and pseudo-random noise information. The method further includes generating, using a machine learning transformer decoder, the action plan based on the input query matrix and a plurality of learnable vectors. The method further includes providing the action plan to the user. The action plan indicates a probability distribution of a plurality of distinct action sequences, to be performed by the user, that transform the start state to the end state.

Probabilistic procedure planning for instructional videos

An apparatus for scrubbing an image may include a memory storing instructions; and a processor configured to execute the instructions to: receive an input image; input a preset public attribute to an encoder neural network; obtain a scrubbed feature from the input image based on the preset public attribute, via the encoder neural network; wherein the encoder neural network is trained based on an amount of information in the scrubbed feature about the input image, and an estimated public attribute estimated from the scrubbed feature.

Method and apparatus for data anonymization

An apparatus for video searching, includes a memory storing instructions, and a processor configured to execute the instructions to split a video into scenes, obtain, from the scenes into which the video is split, one or more textual descriptors describing each of the scenes, encode the obtained one or more textual descriptors describing each of the scenes into a video scene vector of each of the scenes, encode a user query into a query vector having a same semantic representation as that of the video scene vector of each of the scenes into which the one or more textual descriptors describing each of the scenes are encoded, and identify whether the video scene vector of at least one among the scenes corresponds to the query vector into which the user query is encoded.

Apparatus for video searching using multi-modal criteria and method thereof

An image processing system provides the capability of extracting three dimensional motion information from two dimensional images. The image processing system achieves such extraction by analyzing selected image points. The selected image points correspond to the projection of points on a three dimensional object. The object is moving over time (i.e. translating and rotating with respect to the imaging device). The selected image points are processed in parallel by a plurality of processors to first determine the translational velocity. Next, the rotational velocity of the patch is determined. Lastly, the relative depth of each image point is determined in light of the previously determined translational and rotational velocity. The image processing system accomplishes this task by using an imager and a processing means. The system has potential applications in filming and in robotics.

Growing community of inventors

Toronto, Canada

Allan Jepson