Chinese Academy of Sciences

Huawei Technologies Co., Limited

Samsung Electronics Co., Ltd.

Kaiqi Huang

Zhang Zhang

Xin Zhao

Chunfeng Yao

Tieniu Tan

Bailan Feng

Yang Zhou

Zhe Liu

Moon-sik Jeong

Gyu-tae Park

Yueying Kao

Yongzhen Huang

Ran He

Weihua Chen

Dangwei Li

Houjing Huang

Xiaotang Chen

Yupei Wang

Yanhua Cheng

Lijun Cao

An image processing method, a related device, and a computer storage medium are provided. The method includes: obtaining a feature intensity image corresponding to a training image, where an intensity value of a pixel in the feature intensity image is used to indicate importance of the pixel for recognizing the training image, and resolution of the training image is the same as resolution of the feature intensity image; and occluding, based on the feature intensity image, a to-be-occluded region in the training image by using a preset window, to obtain a new image, where the to-be-occluded region includes a to-be-occluded pixel, and the new image is used to update an image recognition model. According to the embodiments of the present application, a prior-art problem that a model has low accuracy and relatively poor generalization performance because of limited training data can be resolved.

Image processing method, related device, and computer storage medium

A method for pedestrian attribute identification and positioning is provided. The method includes: performing feature extraction on a to-be-detected image at a plurality of different abstraction degrees, to obtain a plurality of first feature maps of a pedestrian attribute; performing convolution on the plurality of first feature maps, to obtain a plurality of second feature maps; mapping each second feature map to a plurality of areas (bins) that overlap each other, and performing max pooling on each bin, to obtain a plurality of high-dimensional feature vectors, where the plurality of bins that overlap each other evenly cover each second feature map; processing the plurality of high-dimensional feature vectors into a low-dimensional vector, to obtain an identification result of the pedestrian attribute; and further obtaining a positioning result of the pedestrian attribute based on the plurality of second feature maps and the plurality of high-dimensional feature vectors.

Pedestrian attribute identification and positioning method and convolutional neural network system

A three-dimensional object detection method includes: extracting a target in a two-dimensional image by a pre-trained deep convolutional neural network to obtain a plurality of target objects; determining a point cloud frustum in a corresponding three-dimensional point cloud space based on each target object; segmenting the point cloud in the frustum based on a point cloud segmentation network to obtain a point cloud of interest; and estimating parameters of a 3D box in the point cloud of interest based on a network with the weighted channel features to obtain the parameters of the 3D box for three-dimensional object detection. According to the present invention, the features of the image can be learned more accurately by the deep convolutional neural network and the parameters of the 3D box in the point cloud of interest are estimated based on the network with the weighted channel features.

Three-dimensional object detection method and system based on weighted channel features of a point cloud

A refined segmentation system, method and device of an image shadow area are provided. The system of the present invention includes: a feature extraction network, a reverse fusion network, and a weighted fusion network. The feature extraction network includes a plurality of sampling layers which are arranged sequentially, a plurality of segmentation features of the shadow areas in the input images are obtained through the sampling layers sequentially. The reverse fusion network includes a plurality of layered reverse fusion branches, each of which includes a plurality of feature fusion layers arranged in sequence, and two input features are fused in sequence through each feature fusion layer. The weighted fusion network is configured to perform weighted fusion on outputs of the plurality of reverse fusion branches to obtain a final segmentation result of the shadow area of the input image.

Refined segmentation system, method and device of image shadow area

The present application discloses a method for assessing aesthetic quality of a natural image based on multi-task deep learning. Said method includes: step 1: automatically learning aesthetic and semantic characteristics of the natural image based on multi-task deep learning; step 2: performing aesthetic categorization and semantic recognition to the results of automatic learning based on multi-task deep learning, thereby realizing assessment of aesthetic quality of the natural image. The present application uses semantic information to assist learning of expressions of aesthetic characteristics so as to assess aesthetic quality more effectively, besides, the present application designs various multi-task deep learning network structures so as to effectively use the aesthetic and semantic information for obtaining highly accurate image aesthetic categorization. The present application can be applied to many fields relating to image aesthetic quality assessment, including image retrieval, photography and album management, etc.

Method for assessing aesthetic quality of natural image based on multi-task deep learning

The application discloses a method and an apparatus for recognizing RGB-D objects based on adaptive similarity measure of dense matching item, wherein the method can include at least the following steps: convolution neural network features of a to-be-queried object and a reference object are extracted; dense matching is carried out on the reference object and the to-be-queried object on the basis of the convolution neural network features fused with RGB and depth information; similarity between the reference object and the to-be-queried object is measured according to a dense matching result; and the to-be-queried object is classified based on the similarity between the reference object and the to-be-queried object. With the embodiments of the present application, at least in part, the technical problem of how to improve the robustness of object recognition is solved.

Method and apparatus for recognizing RGB-D objects based on adaptive similarity measure of dense matching item

The present invention relates to a large-range-first cross-camera visual target re-identification method. The method comprises: step S, obtaining initial single-camera tracks of targets; step S, calculating a piecewise major color spectrum histogram feature of each track, and obtaining a track feature representation; step S, obtaining a calculation formula of the similarity between any two tracks by using a minimum uncertainty method, so as to obtain the similarity between any two tracks; and step S, performing global data association on all the tracks by using a maximum posterior probability method, so as to obtain a cross-camera tracking result. The target re-identification method of the present invention achieves high correct identification accuracy.

Large-range-first cross-camera visual target re-identification method

A shape classification method based on the topological perceptual organization (TPO) theory, comprising steps of: extracting boundary points of shapes (S1); constructing topological space and computing the representation of extracted boundary points (S2); extracting global features of shapes from the representation of boundary points in topological space (S3); extracting local features of shapes from the representation of boundary points in Euclidean space (S4); combining global features and local features through adjusting the weight of local features according to the performance of global features (S5); classifying shapes using the combination of global features and local features (S6). The invention is applicable for intelligent video surveillance, e.g., objects classification and scene understanding. The invention can also be used for the automatic driving system wherein robust recognition of traffic signs plays an important role in enhancing the intelligence of the system.

Shape classification method based on the topological perceptual organization theory

Disclosed are an event structure system and a method and medium for controlling the event structure system. The method includes recognizing multiple-person interaction primitives, composing an event by inference based on temporal relations using the multiple-person interaction primitive, and determining a final event by either eliminating an unnecessary event from the composed event, or adding a new event in the composed event.

Growing community of inventors

Beijing, China

Kaiqi Huang