Sony Group Corporation

Sony Corporation

Sony Corporation of America

Ko-Kai Albert Huang

Ming-Chang Liu

Bi Song

Wanxin Xu

Xiaojing Chen

Liyu Gong

Seiji Kobayashi

Tatsumi Sakaguchi

Michael Zordan

Su-Hui Chiang

Joshua Scott Hobson

Implementations generally perform robust multi-view multi-target action recognition using reconstructed 3-dimensional (3D) poses. In some implementations, a method includes obtaining a plurality of videos of a plurality of subjects in an environment, where at least one target subject of the plurality of subjects performs one or more actions in the environment. The method further includes tracking the at least one target subject across at least two cameras. The method further includes reconstructing a 3-dimensional (3D) model of the at least one target subject based on the plurality of videos and the tracking of the at least one target subject. The method further includes recognizing the one or more actions of the at least one target subject based on the reconstructing of the 3D model.

Multi-view multi-target action recognition

Implementations generally recognize clinical activity using multiple cameras. In some implementations, a method includes obtaining a plurality of videos of a plurality of objects in an environment. The method further includes determining one or more key points for each object of the plurality of objects. The method further includes recognizing activity information based on the one or more key points. The method further includes computing workflow information based on the activity information.

Clinical activity recognition with multiple cameras

A method for automatically estimating cellularity in a digital pathology slide image includes: extracting patches of interest from the digital pathology slide image; operating on each patch using a trained first deep convolutional neural network (DCNN) to classify that patch as either normal, having an estimated cellularity of 0%, or suspect, having a cellularity roughly estimated to be greater than 0%; operating on each suspect patch using a second DCNN, trained using a deep ordinal regression model, to determine an estimated cellularity score for that suspect patch; and combining the estimated cellularity scores of the patches of interest to provide an estimated cellularity for the digital pathology slide image at a patch-by-patch level.

Automatic estimation of tumor cellularity using a DPI AI platform

An image-based classification workflow uses unsupervised clustering to help a user identify subpopulations of interest for sorting. Labeled cell images are used to fine-tune the supervised classification network for a specific experiment. The workflow allows the user to select the populations to sort in the same manner for a variety of applications. The supervised classification network is very fast, allowing it to make real-time sort decisions as the cell travels through a device. The workflow is more automated and has fewer user steps, which improves the ease of use. The workflow uses machine learning to avoid human error and bias from manual gating. The workflow does not require the user to be an expert in image processing, thus increasing the ease of use.

Classification workflow for flexible image based particle sorting

A method for super-resolution of block-compressed texture is provided. The method includes receiving a first texture block of a first block size. Based on application of a first block compression (BC) scheme on the received first texture block, coded-texture values are generated in a compressed domain. Further, a first machine learning model is applied on the generated coded-texture values to generate super-resolution coded-texture values in the compressed domain. The generated super-resolution coded-texture values are processed to generate a second texture block of a second block size. The second block size is greater than the first block size.

Super-resolution of block-compressed texture for texture mapping applications

A method of tumor detection and segmentation accepts a first Whole Slide Image (WSI) having a first resolution; creates a corresponding second WSI having a second resolution lower than the first resolution; applies an adaptive thresholding technique to the second WSI to create a background removal mask background; applies the mask to the first WSI to provide a third WSI with extracted patches, characterized by a third resolution, greater than the second resolution and lower than the first resolution; uses a first machine learning system on the third WSI to create a heat map at the third resolution, indicating a subset of the patches likely to include one or more clusters of tumor cells; and uses a second machine learning system on the first WSI and the heat map to segment each patch in a corresponding output image at the first resolution, outlining one or more corresponding clusters.

Tumor detection and segmentation for DPI AI platform routes

Implementations generally perform pose reconstruction by tracking for video analysis. In some implementations, a method includes obtaining a plurality of videos of at least one subject performing at least one action in an environment. The method further includes tracking the at least one subject across at least two cameras. The method further includes reconstructing a 3-dimensional (3D) model of the at least one subject based on the plurality of videos and the tracking of the at least one subject.

Pose reconstruction by tracking for video analysis

Implementations generally relate to surgical scene assessment based on computer vision. In some implementations, a method includes receiving a first image frame of a plurality of image frames associated with a surgical scene. The method further includes detecting one or more objects in the first image frame. The method further includes determining one or more positions corresponding to the one or more objects. The method further includes tracking each position of the one or more objects in other image frames of the plurality of image frames.

Surgical scene assessment based on computer vision

Implementations generally relate to determining the position of a tool tip of a surgical instrument and the orientation of the surgical instrument. In some implementations, a method includes receiving at least one image frame of a plurality of image frames. The method further includes detecting a surgical tool in the at least one image frame. The method further includes determining a position of a tip of the surgical tool. The method further includes determining an orientation of the surgical tool.

Surgical tool tip and orientation determination

A method for automatic organ segmentation in CT images comprises in a first step, rough region segmentation; in a second step, coarse organ segmentation; and in a third step, refinement of organ segmentation. The organ may be a liver. Rough region segmentation may comprise applying standard anatomical knowledge to the CT images. Coarse segmentation may comprise identifying organ voxels using a probabilistic model. Refinement of organ segmentation may comprise refinement based on intensity, followed by refinement based on shape. Apparatuses configured to carry out the method are also disclosed.

Automatic liver segmentation in CT

Implementations generally relate to determining tool handedness for surgical videos. In some implementations, a method includes receiving at least one image frame of a plurality of image frames. The method further includes detecting one or more objects in the at least one image frame. The method further includes classifying the one or more objects into one or more tool classifications, where the one or more objects are tools. The method further includes determining, for each tool, if the tool is assistive or non-assistive. The method further includes determining, for each tool, a handedness of the tool.

Tool handedness determination for surgical videos

Blind legacy video artifact reduction reduces ghost artifacts and Y/C artifacts. Ghost artifacts are reduced using edge correlation and an IIR filter. Y/C artifacts are reduced using frequency domain filtering. The result is much cleaner video content.

Blind legacy video artifact reduction

Implementations generally relate to detecting dominant tools in surgical videos. In some implementations, a method includes receiving at least one image frame. The method further includes detecting one or more objects in the at least one image frame. The method further includes classifying the one or more objects into one or more tool classifications, where the one or more objects are tools. The method further includes determining a handedness of the one or more tools. The method further includes determining a dominant tool from the one or more tools based at least in part on the one or more classifications of the one or more tools and based at least in part on the handedness of the one or more tools.

Dominant tool detection system for surgical videos

A method for 2D feature tracking by cascaded machine learning and visual tracking comprises: applying a machine learning technique (MLT) that accepts as a first MLT input first and second 2D images, the MLT operating on the images to provide initial estimates of a start point for a feature in the first image and a displacement of the feature in the second image relative to the first image; applying a visual tracking technique (VT) that accepts as a first VT input the initial estimates of the start point and the displacement, and that accepts as a second VT input the two 2D images, processing the first and second inputs to provide refined estimates of the start point and the displacement; and displaying the refined estimates in an output image.

Method for 2D feature tracking by cascaded machine learning and visual tracking

An assistive apparatus for organ localization, includes storing a 3D representation and CT images of an anatomical portion of the body of a subject. A localization circuitry determines a rib region and a spine region in the CT images and calculates first and second number of voxels within a first and second region of the 3D representation, respectively. The localization circuitry determines the right side of the body in the CT images, based on a comparison result for the first and second number of voxels. The localization circuitry detects a first bottom portion of right lung based on a distribution of intensity values of pixels in a region of right lung. The localization circuitry detects a second bottom portion of the rib region and localizes the liver organ in the CT images, from a reference of the detected first bottom portion and the detected second bottom portion.

Growing community of inventors

Cupertino, CA, United States of America

Ko-Kai Albert Huang