Baidu USA LLC

Baidu Online Network Technology (beijing) Co., Ltd.

Baidu.com Times Technology (beijing) Co. Ltd.

Beijing Baidu Netcom Science and Technology Co., Ltd.

Baidu Online Network Technology Co., Ltd.

Yingze Bao

Le Kang

Mingyu Chen

Zeyu Liu

Jian Wang

Hao Sun

Errui Ding

Shilei Wen

Xubin Li

Baopu Li

Chengyue Zhang

Xiao Xi Liu

Zhiyu Cheng

Zhizhen Chi

Haofeng Kou

Kuipeng Wang

Yanwen Fan

Qiang Zhou

Hongwu Zhang

Zhihong Pan

Renyi Zhou

Yuexiang Hu

Hsuchun Cheng

Peng Fu

Yanghua Fang

Tao Liu

Song Yang

Wei Qi

Zhigang Wang

Xin Zhou

Yupeng Li

Shengyi He

Qingze Wang

Xuejun Wang

Yuxuan Luo

Hongtian Yang

Jingxi Lu

Haomin Liu

Min-Hung Chen

Xiao Liu

Deep neural networks (DNN) model quantization may be used to reduce storage and computation burdens by decreasing the bit width. Presented herein are novel cursor-based adaptive quantization embodiments. In embodiments, a multiple bits quantization mechanism is formulated as a differentiable architecture search (DAS) process with a continuous cursor that represents a possible quantization bit. In embodiments, the cursor-based DAS adaptively searches for a quantization bit for each layer. The DAS process may be accelerated via an alternative approximate optimization process, which is designed for mixed quantization scheme of a DNN model. In embodiments, a new loss function is used in the search process to simultaneously optimize accuracy and parameter size of the model. In a quantization step, the closest two integers to the cursor may be adopted as the bits to quantize the DNN together to reduce the quantization noise and avoid the local convergence problem.

Cursor-based adaptive quantization for deep neural networks

Described herein are systems and methods that improve the success rate of relocalization and eliminate the ambiguity of false relocalization by exploiting motions of the sensor system. In one or more embodiments, during a relocalization process, a snapshot is taken using one or more visual sensors and a single-shot relocalization in a visual map is implemented to establish candidate hypotheses. In one or more embodiments, the sensors move in the environment, with a movement trajectory tracked, to capture visual representations of the environment in one or more new poses. As the visual sensors move, the relocalization system tracks various estimated localization hypotheses and removes false ones until one winning hypothesis. Once the process is finished, the relocalization system outputs a localization result with respect to the visual map.

Systems and methods for robust self-relocalization in a visual map

Described herein are systems and methods for object detection to achieve hard real-time performance with low latency. Real-time object detection frameworks are disclosed. In one or more embodiments, a framework comprises a first CPU core, a second CPU core, and a plurality of shaves. In one or more embodiments, the first CPU core handles general CPU tasks, while the second CPU core handles the image frames from a camera sensor and computation task scheduling. In one or more embodiments, the scheduled computation tasks are implemented by the plurality of shaves using at least one object-detection model to detect an object in an image frame. In one or more embodiments, computation results from the object-detection model with a higher detection probability is used to form an output for object detection. In one or more embodiments, the object-detection models share some parameters for smaller size and higher implementing speed.

Systems and methods for low-power, real-time object detection

Embodiments of the present disclosure disclose a method and apparatus for compressing a deep learning model. An embodiment of the method includes: acquiring a to-be-compressed deep learning model; pruning each layer of weights of the to-be-compressed deep learning model in units of channels to obtain a compressed deep learning model; and sending the compressed deep learning model to a terminal device, so that the terminal device stores the compressed deep learning model. By pruning each layer of weights of the deep learning model in units of channels, the parameter redundancy of the deep learning model is effectively reduced, thereby improving the computational speed of the deep learning model and maintaining the model accuracy.

Method and apparatus for compressing deep learning model

Described herein are embodiments of a deep residual network dedicated to color filter array mosaic patterns. A mosaic stride convolution layer is introduced to match the mosaic pattern of a multispectral filter arrays (MSFA) or a color filter array raw image. Embodiments of a data augmentation using MSFA shifting and dynamic noise are applied to make the model robust to different noise levels. Embodiments of network optimization criteria may be created by using the noise standard deviation to normalize the Lloss function. Comprehensive experiments demonstrate that embodiments of the disclosed deep residual network outperform the state-of-the-art denoising algorithms in MSFA field.

Deep residual network for color filter array image denoising

A method for information processing for accelerating neural network training. The method includes: acquiring a neural network corresponding to a deep learning task; and performing iterations of iterative training on the neural network based on a training data set. The training data set includes task data corresponding to the deep learning task. The iterative training includes: processing the task data in the training data set using a current neural network, and determining, based on a processing result of the neural network on the task data in a current iterative training, prediction loss of the current iterative training; determining a learning rate and a momentum in the current iterative training; and updating weight parameters of the current neural network by gradient descent based on a preset weight decay, and the learning rate, the momentum, and the prediction loss in the current iterative training. This method achieves efficient and low-cost deep learning-based neural network training.

Method, electronic device and computer readable medium for information processing for accelerating neural network training

Embodiments of the present disclosure provide a method and apparatus for identifying an item. The method includes: acquiring an item image of a to-be-identified item; setting initial position coordinates of the to-be-identified item on the item image; and executing following identifying: inputting the item image and the initial position coordinates into a pre-trained attention module to output an item feature of the to-be-identified item; inputting the item feature into a pre-trained long short-term memory network to output a predicted category and predicted position coordinates of the to-be-identified item; determining whether a preset condition is satisfied; and determining, in response to the preset condition being satisfied, a predicted category of the to-be-identified item outputted by the long short-term memory network a last time for use as a final category of the to-be-identified item.

Method and apparatus for identifying item

The disclosure provides human-tracking methods, systems, and storage media. The method includes: acquiring a plurality of human point clouds of a current frame from a plurality of cameras; generating a total point cloud of the current frame by integrating the plurality of human point clouds of the current frame; acquiring a plurality of human point clouds of a next frame from the plurality of cameras; acquiring a total point cloud of the next frame by integrating the plurality of human point clouds of the next frame; and performing human tracking based on the total point cloud of the current frame and the total point cloud of the next frame.

Human-tracking methods, systems, and storage media

The present disclosure provides a method and a device for recognizing a product, an electronic device and a non-transitory computer readable storage medium, relating to a field of unmanned retail product recognition. The method includes the following. A video taken by each camera in a store is acquired. A recognition is performed on each video to obtain a video segment that a product delivery is recognized and to obtain participated users. The participated users include a delivery initiation user and a delivery reception user. The video segment is inputted into a preset delivery recognition model to obtain a recognition result. The recognition result includes a product delivered and a delivery probability. The product information of products carried by the participated users is updated based on the recognition result.

Method and device for recognizing product

Embodiments disclose an artificial intelligence chip and a convolutional neural network applied to the artificial intelligence chip comprising a processor, at least one parallel computing unit, and a pooling computation unit. The method comprises: dividing a convolution task into convolution subtasks and corresponding pooling subtasks; executing convolution subtasks at different parallel computing units, and performing convolution, batch normalization, and non-linear computing operation in a same parallel computing unit; sending an execution result of each parallel computing unit from executing the convolution subtask to the pooling computation unit for executing the corresponding pooling subtask; merging executing results of the pooling computation unit from performing pooling operations on the executing results outputted by respective convolution subtasks to obtain an execution result of the convolution task. This can reduce data transport, such that operations of the convolutional neural network may be accomplished with lower power consumption and less time in an edge device.

Hardware-software co-design for accelerating deep learning inference

Described herein are systems and embodiments for multispectral image demosaicking using deep panchromatic image guided residual interpolation. Embodiments of a ResNet-based deep learning model are disclosed to reconstruct the full-resolution panchromatic image from multispectral filter array (MSFA) mosaic image. In one or more embodiments, the reconstructed deep panchromatic image (DPI) is deployed as the guide to recover the full-resolution multispectral image using a two-pass guided residual interpolation methodology. Experiment results demonstrate that the disclosed method embodiments outperform some state-of-the-art conventional and deep learning demosaicking methods both qualitatively and quantitatively.

Systems and methods for multispectral image demosaicking using deep panchromatic image guided residual interpolation

The present disclosure provides a human body recognition method and apparatus, and a storage medium, the method comprising: determining a coordinate of a target person in a three-dimensional space according to images containing the target person collected by at least two cameras; calculating back-projection errors of the target person under different cameras respectively according to the coordinate of the target person in the three-dimensional space; determining whether the cameras have a human body recognition error according to the back-projection errors of the cameras; when a camera has the human body recognition error, performing re-recognition of the target person under the camera by using person re-identification ReID, until the back-projection errors of all the cameras containing the target person are not greater than a preset threshold. The present disclosure can improve accuracy of the human body recognition result effectively.

Human body recognition method and apparatus, and storage medium

A human body tracking method, apparatus, and device, and a storage medium. The method includes: obtaining a current frame image captured by a target photographing device at a current moment; detecting each human body in the current frame image to obtain first position information of the each human body in the current frame image; calculating second position information of a first human body in the current frame image; determining target position information of the each human body in the current frame image according to the second position information of the first human body in the current frame image, the first position information of the each human body in the current frame image, and pedestrian features of all tracked pedestrians stored in a preset list.

Human body tracing method, apparatus and device, and storage medium

Embodiments of the present disclosure provide a method and apparatus for generating information, a device for human-computer interaction, and a computer readable medium. The method may include: acquiring gravity sensing data of a shelf carrying an item; and identifying, in response to determining that the item on the shelf is taken based on the gravity sensing data, the taken item based on the gravity sensing data and an acquired image of the taken item, and generating order information of the taken item.

Method and apparatus for generating information, and device for human-computer interaction

A method, apparatus, device, and storage medium for predicting the number of people of a dense crowd, including: converting a first image, in which the number of people is to be determined, into a corresponding first thermodynamic chart according to a thermodynamic chart conversion model; and determining the number of people in the first image according to the first thermodynamic chart, wherein the thermodynamic chart conversion model is obtained by training according to a pre-marked second image and a thermodynamic chart corresponding to each second image, thereby achieving prediction of the number of people of a dense crowd, improving the accuracy in predicting the number of people of the dense crowd while improving management efficiency.

Method, apparatus, device, and storage medium for predicting the number of people of dense crowd

Growing community of inventors

Sunnyvale, CA, United States of America

Yingze Bao