Beijing Sensetime Technology Development Co., Ltd.

Dahua Lin

Xiaoou Tang

Yue Zhao

Kai Chen

Chen Change Loy

Jiaqi Wang

Rui Li Xu

Jianping Shi

Shuo Yang

Zhirong Wu

Yu Qiao

Ziwei Liu

Limin Wang

Zhiyu Min

Yuanjun Xiong

Yule Li

Dian Shao

Yu Xiong

Qingqiu Huang

Methods and apparatuses for deep learning-based recommendation, electronic devices, and media include: respectively obtaining related information of a target user and related information of a target item; respectively using at least two sub-models in an integrated model to obtain, based on the related information of the target user and the related information of the target item, operating probabilities corresponding to the at least two sub-models; obtaining, based on the operating probabilities corresponding to the at least two sub-models, a target probability about the target user operating the target item; and recommending the target item to the target user based on the target probability.

Methods and apparatuses for deep learning-based recommendation, electronic devices, and media

An anchor determination includes: performing feature extraction on an image to be processed to obtain a first feature map of the image to be processed; performing anchor prediction on the first feature map via an anchor prediction network to obtain position information of anchors and shape information of the anchors in the first feature map, the position information of the anchors referring to information about positions in the first feature map where the anchors are generated. A corresponding anchor determination apparatus and a storage medium are also provided.

Anchor determination method and apparatus, electronic device, and storage medium

A key frame scheduling method and apparatus include: performing feature extraction on a current frame through a first network layer of a neural network to obtain low-layer features of the current frame acquiring a scheduling probability of the current frame according to low-level features of a previous key frame adjacent to the current frame and the low-level features of the current frame; determining whether the current frame is scheduled as a key frame according to the scheduling probability value of the current frame; and when determining that the current frame is scheduled as a key frame, performing feature extraction on low-level features of a current key frame via a second network layer of the neural network to obtain high-level features of the current key frame, where the network depth of the first network layer is less than the network depth of the second network layer.

Keyframe scheduling method and apparatus, electronic device, program and medium

The present disclosure relates to a method and device for video processing, an electronic device, and a storage medium. The method comprises: determining, on the basis of paragraph information of a query text paragraph and video information of multiple videos in a video library, preselected videos associated with the query text paragraph in the multiple videos; and determining a target video in the preselected videos on the basis of video frame information of the preselected videos and of sentence information of the query text paragraph. The method for video processing of the embodiments of the present disclosure indexes videos by means of the relevance between the videos and the query text paragraph, allows the pinpointing of the target video, avoids search result redundancy, allows the processing of the query text paragraph in a natural language form, and is not limited by the inherent contents of content labels.

Method and device for video processing, electronic device, and storage medium

An example of the present disclosure provides methods, apparatuses and devices for magnifying a feature map, and a computer readable storage medium. The method includes: receiving a source feature map to be magnified; obtaining N reassembly kernels corresponding to each source position in the source feature map by performing convolution on the source feature map, wherein N refers to a square of a magnification factor of the source feature map; obtaining, for each of the reassembly kernels, a normalized reassembly kernel by performing normalization; obtaining, for each source position in the source feature map, N reassembly features corresponding to the source position by reassembling features of a reassembly region determined according to the source position with N normalized reassembly kernels corresponding to the source position; and generating a target feature map according to the N reassembly features corresponding to each source position in the source feature map.

Magnifying feature map

Time domain action detecting methods and systems, electronic devices, and computer storage medium are provided. The method includes: obtaining a time domain interval in a video with an action instance and at least one adjacent segment in the time domain interval; separately extracting action features of at least two video segments in candidate segments, where the candidate segments comprises video segment corresponding to the time domain interval and adjacent segments thereof; pooling the action features of the at least two video segments in the candidate segments, to obtain a global feature of the video segment corresponding to the time domain interval; and determining, based on the global feature, an action integrity score of the video segment corresponding to the time domain interval. The embodiments of the present disclosure benefit accurately determining whether a time domain interval comprises an integral action instance, and improve the accuracy rate of action integrity identification.

Time domain action detecting methods and system, electronic devices, and computer storage medium

Growing community of inventors

Beijing, China

Dahua Lin