Intel Corporation

Samsung Electronics Co., Ltd.

Haibing Ren

Yimin Zhang

Ping Guo

Jiali Zhao

Wei Hu

Fei Duan

Seokcheol Kee

Haitao Wang

Lidan Zhang

Xiaobo Hu

Jong Ha Lee

Hu Tiger Chen

Wonjun Hwang

Jung Bae Kim

Hwa Sup Lim

Dejun Wang

Seok-cheol Kee

Jiankun Hu

Zhongxuan Liu

Xiangsheng Huang

Sirui Yang

A long-term object tracker employs a continuous learning framework to overcome drift in the tracking position of a tracked object. The continuous learning framework consists of a continuous learning module that accumulates samples of the tracked object to improve the accuracy of object tracking over extended periods of time. The continuous learning module can include a sample pre-processor to refine a location of a candidate object found during object tracking, and a cropper to crop a portion of a frame containing a tracked object as a sample and to insert the sample into a continuous learning database to support future tracking.

Continuous learning for object tracking

A system, article, and method of neural network object recognition for image processing includes customizing a training database and adapting an instance segmentation neural network used to perform the customization.

Method and system of neural network object recognition for image processing

A system and method of teaching a neural network through reinforcement learning methodology. The system includes a machine-readable medium having one or more processors that perform a motion task to produce a first result corresponding to navigating a device during a first episode and performing an interaction task during that same episode. After completion of the first episode a processor calculates a Q value change based on the first task result and the second task result. The processor then modifies parameters based on the Q value change such that during subsequent episode iterations the motion task and interactive task are improved and a smooth and continuous transition occurs between these two tasks.

Reinforcement learning for human robot interaction

A system, article, and method of point cloud registration using overlap regions for image processing.

Method and system of point cloud registration for image processing

Methods and apparatus to match images using semantic features are disclosed. An example apparatus includes a semantic labeler to determine a semantic label for each of a first set of points of a first image and each of a second set of points of a second image; a binary robust independent element features (BRIEF) determiner to determine semantic BRIEF descriptors for a first subset of the first set of points and a second subset of the second set of points based on the semantic labels; and a point matcher to match first points of the first subset of points to second points of the second subset of points based on the semantic BRIEF descriptors.

Methods and apparatus to match images using semantic features

Techniques are provided for estimation of human orientation and facial pose, in images that include depth information. A methodology embodying the techniques includes detecting a human in an image generated by a depth camera and estimating an orientation category associated with the detected human. The estimation is based on application of a random forest classifier, with leaf node template matching, to the image. The orientation category defines a range of angular offsets relative to an angle corresponding to the human facing the depth camera. The method also includes performing a three dimensional (3D) facial pose estimation of the detected human, based on detected facial landmarks, in response to a determination that the estimated orientation category includes the angle corresponding to the human facing the depth camera.

Estimation of human orientation in images using depth information from a depth camera

Techniques are provided for detection of humans in images that include depth information. A methodology embodying the techniques includes segmenting an image into multiple windows and estimating the distance to a subject in each window based on depth pixel values in that window, and filtering to reject windows with sizes that are outside of a desired window size range. The desired window size range is based on the estimated subject distance and the focal length of the depth camera that produced the image. The method further includes generating classifier features for each remaining windows (post-filtering) for use by a cascade classifier. The cascade classifier creates candidate windows for further consideration based on a preliminary detection of a human in any of the remaining windows. The method further includes merging neighboring candidate windows and executing a linear classifier on the merged candidate windows to verify the detection of a human.

Detection of humans in images using depth information

The present disclosure is directed to face detection window refinement using depth. Existing face detection systems may perform face detection by analyzing portions of visual data such as an image, video, etc. identified by sub-windows. These sub-windows are now determined only based on pixels, and thus may number in the millions. Consistent with the present disclosure, at least depth data may be utilized to refine the size and appropriateness of sub-windows that identify portions of the visual data to analyze during face detection, which may substantially reduce the number of sub-windows to be analyzed, the total data processing burden, etc. For example, at least one device may comprise user interface circuitry including capture circuitry to capture both visual data and depth data. Face detection circuitry in the at least one device may refine face detection by determining criteria for configuring the sub-windows that will be used in face detection.

Face detection window refinement using depth

Systems, apparatuses and methods may generate a map of a search environment based on a probability of a target human being present within the search environment, capture a red, green, blue, depth (RGBD) image of one or more potential target humans in the search environment based on the map, and cause a robot apparatus to obtain a frontal view position with respect to at least one of the one or more potential target humans based on the RGBD images.

Human search and identification in complex scenarios

Disclosed is a 3D face capturing apparatus, method and computer-readable medium. As an example, the 3D face capturing method includes obtaining a face color image, obtaining a face depth image, aligning, by a computer, the face color image and the face depth image, obtaining, by the computer, a 3D face model by 2D modeling of the face color image and covering a modeled 2D face area on an image output by an image alignment module, removing by the computer, depth noise of the 3D face model, and obtaining, by the computer, an accurate 3D face model by aligning the 3D face model and a 3D face template, and removing residual noise based on a registration between the 3D face model and the 3D face template.

Three-dimensional face capturing apparatus and method and computer-readable medium thereof

Provided are a detector and a method of detecting an object using the detector. The method includes combining a first detector and a second detector in a combination scheme to form a multi-layer combination detector, the second detector being of a type different from that of the first detector, processing a binary classification detection with respect to an inputted sample starting from an uppermost layer detector, allowing a sample of an object detected from a current layer to approach a lower layer, while rejecting a sample of a non-object detected from the current layer whereby the rejected non-object may not approach the lower layer, and outputting a sample passing through all layers as a detected object.

Combination detector and object detection method using the same

Provided are an object verification apparatus and method. The object verification apparatus includes a matching unit performing a plurality of different matching algorithms on a query image and generating a plurality of scores; a score normalization unit normalizing each of the generated scores to be adaptive to the query image; a weight estimation unit estimating weights of the normalized scores based on the respective matching algorithms applied; and a score fusion unit fusing the normalized scores by respectively applying the weights estimated by the weight estimation unit to the normalized scores.

Object verification apparatus and method

An apparatus for and method of performing a most informative feature extraction (MIFE) method in which a facial image is separated into sub-regions, and each sub-region makes individual contribution for performing facial recognition. Specifically, each sub-region is subjected to a sub-region based adaptive gamma (SadaGamma) correction or sub-region based histogram equalization (SHE) in order to account for different illuminations and expressions. A set of reference images is also divided into sub-regions and subjected to the SadaGamma correction or SHE. A comparison is made between the each corrected sub-region and each corresponding sub-region of the reference images. Based upon the comparisons made individually for the sub-regions of the facial image, one of the stored reference images having the greatest correspondence is chosen. While usable individually, using the MIFE and/or SadaGamma correction or SHE together achieves a lower error ratio in face recognition under different expressions, illuminations and occlusions.

Apparatus for and method of feature extraction for image recognition

There is provided a method of estimating disparity for 3D object recognition. The method includes obtaining a plurality of images having different resolutions for the stereo image, estimating a disparity map for a lowest-resolution level image, estimating a coarse disparity map for an upper resolution level image by using the disparity, obtaining a fine disparity map for the upper resolution level image by using the coarse disparity, and outputting the fine disparity map as a final disparity map for the stereo image if the upper resolution level image has a resolution of a highest level.

Method of estimating disparity for 3D object recognition

A image verification method, medium, and apparatus using a local binary pattern (LBP) discriminant technique. The verification method includes generating a kernel fisher discriminant analysis (KFDA) basis vector by using the LBP feature of an input image, obtaining a Chi square inner product by using the LBP feature of an image registered in advance and a kernel LBP feature and projecting to a KFDA basis vector, obtaining a Chi square inner product by using the LBP feature of a query image and a kernel LBP feature and projecting to a KFDA basis vector, and obtaining the similarity degree of the target image and the query image that are obtained as Chi square inner product results, and projected to the KFDA basis vector. According to the method, medium, and apparatus, the KFDA based LBP shows superior performance over conventional LBP, KFDA, and biometric experimentation environment (BEE) baseline algorithms.

Growing community of inventors

Beijing, China

Haibing Ren