Abstract: Tensor decomposition can be advantageous for compressing deep neural networks (DNNs). In many applications of DNNs, reducing the number of parameters and computation workload is helpful to accelerate inference speed in deployment. Modern DNNs comprise multiple layers with multi-array weights where tensor decomposition is a natural way to perform compression—in which the weight tensors in convolutional layers or fully-connected layers are decomposed with specified tensor ranks (e.g., canonical ranks, tensor train ranks). Conventional tensor decomposition with DNNs involves selecting ranks manually, which requires tedious human efforts to finetune the performance. Accordingly, presented herein are rank selection embodiments, which are inspired by reinforcement learning, to automatically select ranks in tensor decomposition.

Abstract: Various techniques for theft proofing autonomous driving vehicles (ADV) for transporting goods are described. In one embodiment, sensor data of a moving object representing a person within a predetermined proximity of an ADV are captured for real-time analysis by a theft detection module, to determine a moving behavior of the moving object based on the sensor data in view of a set of known moving behaviors. The theft detection module further determines whether an intention of the person is likely to remove at least some of the goods from the ADV using a process derived from historical image set, and sends an alarm to a predetermined destination in response to determining such an intention of the person. Other sensor data, for example, real time movements and weights of the ADV, can be used in conjunction with the process derived from historical image sets to determine the intention of the person.

Abstract: 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.

Abstract: A first request is received from a first processing node to produce data blocks of a first data stream representing a first communication topic. The first processing node is one of the processing nodes handling a specific function. Each of the processing nodes is executed within a specific node container having a specific operating environment. A global memory segment is allocated from a global memory to store the data blocks of the first data stream. A first local memory segment is mapped to the global memory segment. The first local memory segment is allocated from a first local memory of a first node container containing the first processing node. The first processing node directly accesses the data blocks of the first data stream stored in the global memory segment by accessing the mapped first local memory segment within the first node container.

Abstract: A first localization system performs a first localization using a first set of sensors to track locations of the ADV along the path from a starting point to a destination point. A first localization curve is generated as a result representing the locations of the ADV along the path tracked by the first localization system. Currently, a second localization system performs a second localization using a second set of sensors to track the locations of the ADV along the path. A second localization curve is generated as a result representing the locations of the ADV along the path tracked by the second localization system. A system delay of the second localization system is determined by comparing the second localization curve against the first localization curve as a localization reference. The system delay of the second localization system can then be utilized to compensate path planning of the ADV subsequently.

Abstract: Presented are systems and methods for improving speed and quality of real-time per-pixel depth estimation of scene layouts from a single image by using an end-to-end Convolutional Spatial Propagation Network (CSPN). An efficient linear propagation model performs propagation using a recurrent convolutional operation. The affinity among neighboring pixels may be learned through a deep convolutional neural network (CNN). The CSPN may be applied to two depth estimation tasks, given a single image: (1) to refine the depth output of existing methods, and (2) to convert sparse depth samples to a dense depth map, e.g., by embedding the depth samples within the propagation procedure. The conversion ensures that the sparse input depth values are preserved in the final depth map and runs in real-time and is, thus, well suited for robotics and autonomous driving applications, where sparse but accurate depth measurements, e.g., from LiDAR, can be fused with image data.

Abstract: According to one embodiment, a system performs a secure boot using a security module such as a trusted platform module (TPM) of a host system. The system establishes a trusted execution environment (TEE) associated with one or more processors of the host system. The system launches a memory manager within the TEE, where the memory manager is configured to manage memory resources of a data processing (DP) accelerator coupled to the host system over a bus, including maintaining memory usage information of global memory of the DP accelerator. In response to a request received from an application running within the TEE for accessing a memory location of the DP accelerator, the system allows or denies the request based on the memory usage information.

Abstract: 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.

Abstract: Presented are deep learning-based systems and methods for fusing sensor data, such as camera images, motion sensors (GPS/IMU), and a 3D semantic map to achieve robustness, real-time performance, and accuracy of camera localization and scene parsing useful for applications such as robotic navigation and augment reality. In embodiments, a unified framework accomplishes this by jointly using camera poses and scene semantics in training and testing. To evaluate the presented methods and systems, embodiments use a novel dataset that is created from real scenes and comprises dense 3D semantically labeled point clouds, ground truth camera poses obtained from high-accuracy motion sensors, and pixel-level semantic labels of video camera images. As demonstrated by experimental results, the presented systems and methods are mutually beneficial for both camera poses and scene semantics.

Abstract: In one embodiment, a system is designed to determine the requirement of a perception range for a particular type of vehicles and a particular planning and control technology. A shadow filter is used to connect a scenario based simulator and a PnC module, and tuning the parameters (e.g. decreasing the filter range, tuning the probability of obstacles to be observed among frames) of shadow filter to mimic the real world perceptions with a limited range and reliabilities. Based on the simulation results (e.g., a failure rate, smoothness, etc.), the system is able to determine the required perception distance for the current PnC module. A PnC module represents a particular autonomous driving planning and control technology for a particular type of autonomous driving vehicles. Notice that the PnC module is replaceable so that this method is suitable for different PnC algorithms representing different autonomous driving technologies.

Abstract: Embodiments of the disclosure disclose a method and an apparatus for calibrating a vehicle control parameter, an on-board controller, and an autonomous vehicle; one embodiment of the method comprises: executing a calibrating step in response to reaching a preset update condition, the calibrating step comprises: obtaining a current offset data set, wherein the current offset data in the current offset data set are determined in a period of time including a current time point; determining a current offset data reference value for characterizing a value feature of the current offset data set; and performing offset correction for the vehicle control parameter based on an offset between the current offset data reference value and a historical offset data reference value. This embodiment implements autonomous calibration of the vehicle parameter based on changes of vehicle offset, such that the vehicle may accurately follow a corresponding control indicator.

Abstract: Embodiments of the present disclosure disclose a method for determining a vehicle control parameter, an apparatus for the same, a vehicle controller, and an autonomous vehicle. An embodiment of the method comprises: determining a current vehicle speed and an expected acceleration; determining, from a pre-generated parameter calibration table, a longitudinal control parameter corresponding to the current vehicle speed and the expected acceleration; wherein the parameter calibration table is obtained by: obtaining a training sample set, a training sample in the training sample set including the vehicle speed, the longitudinal control parameter and the acceleration; training a pre-established parameter calibration model with the vehicle speed and acceleration of respective samples in the training sample set as inputs and the longitudinal control parameter value in the training sample as an expected output; and obtaining the parameter calibration table based on the trained parameter calibration model.

Abstract: Embodiments of the present disclosure provide a method and apparatus for generating a video. The method may include: receiving a query text inputted by a user; querying a material resource set related to the query text, material resources being images, videos, or audios; presenting the material resource set; determining a material resource sequence, in response to receiving a selecting operation and a ranking operation of the user on the material resources in the presented material resource set; and generating the video based on the material resource sequence.

Abstract: Embodiments of the present disclosure provide a method and apparatus for generating a video. The method may include: determining a commentary of a target news cluster, each piece of news in the target news cluster being specific to a given news event; generating a voice corresponding to each paragraph in the commentary using a speech synthesis technology; determining a candidate material resource set corresponding to the commentary based on a video and an image included in the target news cluster, the candidate material resource being a video or image; determining a candidate material resource sequence corresponding to the each paragraph in the commentary; and generating a video corresponding to the commentary based on the voice corresponding to the each paragraph in the commentary and the candidate material resource sequence.

Abstract: Embodiments of the present disclosure provide a method and apparatus for generating information. The method may include: determining at least one video segment obtained by semantically segmenting videos included in a target news cluster as a target video set, where respective pieces of news in the target news cluster directs to a given news event; determining a commentary for the target news cluster; determining, based on the target video set and a target image set, a candidate material resource set corresponding to the commentary, where the target image set is composed of respective images included in the target news cluster; and for each paragraph in the commentary, determining degrees of matching between the paragraph and candidate material resources in the candidate material resource set.

Abstract: Presented are deep learning-based systems and methods for fusing sensor data, such as camera images, motion sensors (GPS/IMU), and a 3D semantic map to achieve robustness, real-time performance, and accuracy of camera localization and scene parsing useful for applications such as robotic navigation and augment reality. In embodiments, a unified framework accomplishes this by jointly using camera poses and scene semantics in training and testing. To evaluate the presented methods and systems, embodiments use a novel dataset that is created from real scenes and comprises dense 3D semantically labeled point clouds, ground truth camera poses obtained from high-accuracy motion sensors, and pixel-level semantic labels of video camera images. As demonstrated by experimental results, the presented systems and methods are mutually beneficial for both camera poses and scene semantics.

Abstract: In one embodiment, in response to a route from a source location to a target location, the route is analyzed to identify a list of one or more driving scenarios along the route that match one or more predetermined driving scenarios. The route is segmented into a list of route segments based on the driving scenarios. At least one of the route segments corresponds to one of the identified driving scenarios. A path is generated based on the route segments for driving an autonomous driving vehicle from the source location to the target location. The path includes a number of path segments corresponding to the route segments. At least one of the path segments of the path is determined based on a preconfigured path segment of a predetermined driving scenario associated with the path segment, without having to calculating the same at real time.

Abstract: 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.

Abstract: A method and apparatus for adjusting a parameter are provided. The method may include: acquiring a current value of at least one parameter which is in a process of generating a video corresponding to a commentary of the news cluster based on a news cluster; determining a video evaluation score of the video which is generated based on the news cluster and according to the current value of the at least one parameter; performing feature extraction on the current value of the at least one parameter to obtain a feature representation; inputting the feature representation and the determined video evaluation score into a pre-trained evaluation network to obtain a predicted video evaluation score; inputting the feature representation and the predicted video evaluation score into a pre-trained operation network, to obtain current operation information; and adjusting the current value of the at least one parameter based on the current operation information.

Abstract: A data center cooling system includes a housing to contain electronic racks of IT components operating therein, a coolant distribution unit (CDU) situated within the housing to control a liquid flow of a cooling liquid. One or more liquid cooling devices are disposed on the IT components to receive a first liquid from the CDU, to exchange or extract heat generated from the IT components, to transform the first liquid to a second liquid with a higher temperature, and to transmit the second liquid back to the CDU. The CDU is coupled to a heat transfer system to dissipate the exchanged heat to an external environment. The data center cooling system includes an airflow delivery system to generate a direct or indirect airflow to travel through the servers of the electronic racks to remove heat generated by the servers to the external environment eliminating chiller and IT fans.