Search

3D object detection is a crucial research topic in computer vision, which usually uses 3D point clouds as input in conventional setups. Recently, there is a trend of leveraging multiple sources of input data, such as complementing the 3D point cloud with 2D images that often have richer color and fewer noises. However, due to the heterogeneous geometrics of the 2D and 3D representations, it prevents us from applying off-the-shelf neural networks to achieve multimodal fusion. To that end, we propose Bridged Transformer (BrT), an end-to-end architecture for 3D object detection. BrT is simple and effective, which learns to identify 3D and 2D object bounding boxes from both points and image patches. A key element of BrT lies in the utilization of object queries for bridging 3D and 2D spaces, which unifies different sources of data representations in Transformer. We adopt a form of feature aggregation realized by point-to-patch projections which further strengthen the correlations between images and points. Moreover, BrT works seamlessly for fusing the point cloud with multi-view images. We experimentally show that BrT surpasses state-of-the-art methods on SUN RGB-D and ScanNetV2 datasets., Comment: CVPR 2022

Land Use and Land Cover (LULC) are the basic units of human activities and also serve as the significant factors to assess the climate change studies and environmental protection. Therefore, it is of significance to accurately and timely obtain the LULC maps on the earth’s surface, in particular for those regions where dramatic LULC changes are undergoing. Since 2017 April, a new state-level new area, Xiong’an New Area, was established in China, which would inevitably lead to some LULC changes in this region. In order to better understand what kinds of LULC change in this region with more details and further evaluate the potential impacts that the LULC changes will bring, this study makes full use of the 10-m multi-temporal Sentinel-2 images on the cloud-computing Google Earth Engine (GEE) and the powerful classification capability of Random Forest (RF) models, to generate the continuous LULC maps in Xiong’an New Area from 2017 to 2020. The derived LULC map for each year in Xiong’an New Area achieves high accuracy, with OA and Kappa values less than 0.9. Based on the obtained LULC maps, this study analyzed the spatial and temporal changes of LULC types in the last four years. It revealed that the estimated dry farmland has decreased from 2017 to 2018 and then kept almost stable, and the majority of it has converted to non-cropland, especially to the tree seedlings and landscape dual-use regions. The estimated impervious areas that were mainly composed of buildings and transport infrastructures first increased due to the newly constructed Xiong’an Station with some high-speed roads and then decreased from 2019 to 2020, mainly due to the relocation of some small villages. Accordingly, some building groups and aggregations were going to show in this region, and are expected to further developed in the future. Other LULC types such as lakes and forests have not changed dramatically, indicating that the environmental protection in Xiong’an New Area was, to date, satisfactory. The obtained 10-m and 4-year LULC maps in this study can provide some valuable information on the monitoring and understanding of what kinds of changes were and will be undergoing in Xiong’an New Area, and can also be used for evaluating the potential impacts and challenges such as environmental protection. Additionally, the utilization of GEE and the multi-temporal 10-m Sentinel-2 images can help to achieve LULC mapping in this region in an accurate and timely manner, which can be used in future studies.

Classification of staple crops is of significance to acquire the updating information on crop distribution, crop yield and related market storage. In the existing studies, the open accessible satellite images are mostly used, especially for MODIS and Landsat images. However, the coarser spatial resolution cannot accurately depict the crop spatial distributions. Additionally, the classification accuracy is often carried out through some randomly selected pixels or polygons, which might not fully evaluate the classification results. In view of this, this study utilizes the Gaofen-1 and Gaofen-6 satellite images with a spatial resolution of 16 m, to extract crop distributions, which is rarely explored in the literature. Several Experiments are conducted in Yanshou county, where a full-scale field investigation is carried out to collect the validation sample data. Through this rigorous assessment, the crop classification maps are thoroughly evaluated in terms of accuracy and running cost. The results indicate that Gaofen-1 and Gaofen-6 images are qualified to accurately classify the staple crops and the splitting tree number variable for Random Forest model can be set to some mid-value such as 100 on consideration of both accuracy and computational efficiency.

Since more and more distributed generators (DGs) are connected into the low voltage distribution network, the existing fault-location methods are difficult to achieve accurate fault location due to its decentralized power supply characteristics. Aiming at this problem and the shortcomings of the existing traveling-wave fault-location methods, this paper proposes an accurate fault-location method for distribution networks with DGs based on multi-terminal traveling wave. Firstly, this paper analyzes the impact of DGs on the first arrival times of the fault-generated traveling waves. Secondly, a set of fault distance data independent of traveling-wave velocity are calculated by using the traveling-wave time information of all detectors, and the faulted branch is identified based on the discrete degree of this set of data. Finally, the quartile method is used to reject the bad data, and the accurate fault distance of the fault point is obtained. Simulation results show that the proposed method can accurately locate the fault point, hardly impacted by traveling-wave velocity, and has good robustness to arrival time errors.

Due to the high concentration of electricity generation and distribution in the floating nuclear power plant grid-source direct connection system, the generator stator ground fault protection is affected by the direct connection power grid. Traditional protection methods cannot distinguish the fault direction and are difficult to guarantee the extreme safety requirements of the nuclear islands. This paper analyzes the system ground fault, points out the limitations of the existing generator stator ground protection methods, and proposes an adaptive form of generator stator ground fault protection method. Based on the phase-current differential, the method can effectively distinguish the faults inside or outside the generator, which can also form the protection criterion and action strategies according to the fault degree. It can trip the generator smoothly under slight fault conditions, ensuring the safe and stable operation of the floating nuclear power plant. Simulation in PSCAD/EMTDC proves that the proposed method can effectively realize the selective coordination of the ground fault protection for the grid-source direct connection system.

In this paper, we propose an embarrassingly simple approach for one-shot learning. Our insight is that the one-shot tasks have domain gap to the network pretrained tasks and thus some features from the pretrained network are not relevant, or harmful to the specific one-shot task. Therefore, we propose to directly prune the features from the pretrained network for a specific one-shot task rather than update it via an optimized scheme with complex network structure. Without bells and whistles, our simple yet effective method achieves leading performances on miniImageNet (60.63%) and tieredImageNet (69.02%) for 5-way one-shot setting. The best trial can hit to 66.83% on miniImageNet and 74.04% on tieredImageNet, establishing a new state-of-the-art. We strongly advocate that our method can serve as a strong baseline for one-shot learning. The codes and trained models will be released at http://github.com/corwinliu9669/embarrassingly-simple-baseline.

In this paper, we address the problem of space-time adaptive processing (STAP) for airborne collocated multiple-input multiple-output(MIMO) radars using a generalized sidelobe canceller(GSC). A collocated MIMO radar that transmits orthogonal waveforms is equivalent to a virtual array with coherent amplitude weights. On this account, the blocking matrix(BM) of the GSC is not orthogonal to the target steering vector and the output SCNR of the GSC processor degrades. In order to improve the performance, the modified Gram-Schmidt orthogonalization(MGSO)-based channel preprocessing method is proposed to modify the blocking matrix of the GSC processor. We prove that this modified GSC processor can achieve the optimal STAP performance. We also show that the proposed method has reduced computational complexity and fast convergence rate. Numerical simulations validate the effectiveness of the proposed methods.

An adaptive-data-prediction-based (ADPB) ultrasound receiver is developed to achieve a wide dynamic range (DR) and high signal-to-noise ratio (SNR) ultrasound imaging with no extra power penalty. To guarantee both high resolution and power efficiency, a digital linear predictor is introduced, which significantly enhances the resolution and eliminates the power hungry circuits in traditional receivers. With digital assisted signal processing, performance stress on analog modules including ADCs is relaxed, resulting in further reduced power consumption. To enable real-time digital signal processing, least-mean-square (LMS) algorithm is introduced. The proposed system was implemented on a FPGA-based hardware platform. The measurement results show that with a simple 10-bit ADC, it achieves 80-dB SNR, which can be usually accomplished by a state-of-art 16-bits ADC. More importantly, the elimination of TGC and the employment of low resolution ADC reduce the total power by 2–3 orders.

The maximum space-time degrees of freedom (DOF) is restricted to the number of antennas and pulses. In this paper, a novel sparse space time adaptive processing (STAP) scheme is proposed based on the concept of minimum redundancy arrays.We arrange the array geometry and the temporal sampler interval to make the location of joint space-time samples satisfies minimum redundancy interval. According to the structural information imposed on clutter covariance matrix (CCM), the original CCM is estimated by a series of matrix basis accurately, and a higher dimension of CCM can be reconstructed while the virtual spatial and temporal elements are adjustable. Simulation results verify the proposed formulations and the angle-Doppler resolution is increased significantly.

The paper presents the average signal-to-clutter loss (SCRL) analysis for polarimetric space-time adaptive processing by exploiting the Kronecker structure of the clutter covariance matrix (CM). An expression for the average SCRL as a function of the mean square error of the corresponding CM estimator is derived. Based on that expression, one can determine how many samples are required in order to achieve a desired SCRL. The proposed average SCRL analysis methodology can be extended to more general scenarios, where closedform CM estimates are not available. Simulations indicate that even in the non-asymptotic regime, the proposed method can provide a good prediction of the average SCRL.

A robust colored-loading factor optimization approach referred to as LWEC is newly proposed for colored-loading knowledge-aided STAP. The colored-loading factor is optimized to maximize the whitening ability of the colored-loaded clutter-and-noise covariance matrix (CNCM) of the cell under test (CUT) on selected secondary samples. The proposed approach is robust to accuracy degree heterogeneity of the a priori information in both the range domain and doppler domain for its ability to evaluate the performance of the colored-loaded CNCM of each range bin independently and conduct the optimization algorithm after the dimension-reduced process. Solution of the optimization function is proved to be unique. Numerical experiment results demonstrate the effectiveness and robustness of the proposed approach.

This study proposes a flexible method to track and quantify the ac-coupled gain variations in electrode to tissue interface (ETI) of adhesive/reusable electrodes. Particularly, this study focuses on the effects of multiplicative motion artifacts (MMA) on ETIs and proposes a continuously monitoring technique to assess ETI multiplicative variations. The proposed method only requires one additional channel of self electrocardiogram (self-ECG). Both gel-covered and foam-based adhesive electrodes were tested. Empirical results show that the ETI multiplicative variations can be characterized by the proposed convenient method. Meanwhile, we successfully demonstrate different characteristics for two types of electrodes under five motion postures including suppress motion, fist/stretch palm, twist, bent inward and arm lift, Furthermore, The long-term degradation experiment illustrates that the average characteristics variation of ETI is less than 6% within 3-day of use.

Recently a few works of semi-supervised learning methods based on graph have been proposed for remote sensing. The common idea of these methods are that they build a graph using the samples of the image. Most of their time complexity is relatively large, and they ignore the spatial information of the image, which leads to unsatisfactory classification results. this paper proposes a novel semi-supervised classification method based on anchor-spatial graph for large PolSAR data. Firstly the unsupervised Wishart clustering is performed to select representative samples, which served as anchors according to the least distance between samples. Then an anchor graph is built using the selected anchors according to the multiple features of the samples. And it is further combined with the spatial information of the samples to construct an anchor-spatial graph. Finally the class information from small quantities of labeled samples propagates to the unlabeled ones. Experimental results show that the proposed method has a low time complexity compared with existing works and it could effectively cut down the processing time for large PolSAR data meanwhile keeps the classification accuracy.

Lip detection is the major and important step in the lip-reading system. In this paper, a novel method is proposed. Firstly, since the combination of YCbCr and HSV color model has a good segmentation power on skin color, we used this scheme to avoid the interference of background in the original images. Then, we use Hough transform to locate the eyeballs which are close to the circular after the face is detected. In the third step, a new approach is proposed to extract the lips based on the transformation of HSV color model. More specifically, the width of lip area is obtained by the distribution relationship between eyes and mouth. For the height of lip, threshold processing and gray-level integration projection are analyzed to extract the whole lip area, whether the mouth is open or closed. Extensive experiments show that the proposed method can rapidly and accurately extract the lip area, without the influence of lip shapes, light conditions and backgrounds. Therefore, this procedure can be effectively integrated into the lip-reading system.

Defocus map estimation plays an important role in computer vision and computer graphics applications. For many defocus map estimation methods from a single image, the edge detection is very important for depth map estimation. In this paper, we show a novel method for defocus map estimation from a single image, and give a theoretical model of re-blurred. The method is based on the difference value (D-value) of Gaussian gradient. Simultaneously, the influence of different edge type on defocus map estimating was analyzed. Compared with the Zhuo's[1] method, the results show that our method could work better on the evaluation of the texture sharpness.

To reduce the mutual coupling across the collocated sensor, a distributed array geometry composed of crossed dipole pair is introduced in this paper. Based on the non-collocated vector sensor array, an ESPRIT-based method is newly proposed to estimate the direction of arrival (DOA) and polarization parameter. Due to the special structure of non-collocated array, the DOA and polarization parameter are paired automatically, and thus the computational complexity is greatly reduced. The Cramer-Rao Lower Bound (CRLB) has been derived and the performance of DOA and polarization estimation for the distributed array is also analyzed. Simulation results verify the effective parameter estimation performance of the proposed method by using a spatially non-collocated array.

The passive polarimetric detection problem for vector-sensor array is formulated based on quaternion. The quaternion-formed detectors are deduced and analyzed in the presence of different quaternion proper noise. The properness of quaternion Gaussian noise, especially for the C-proper and Q-proper cases in this paper, is discussed based on the second order statistics (SOS) of the complex components of quaternion noise. Numerical simulations and statistical analysis demonstrate that the C-proper, Q-proper and complex detectors are equivalent in the Q-proper case, and that the Q-proper detector behaves poorer than C-proper and complex detectors in the C-proper case.

Quaternion technique is applied to model the polarimetric statistical MIMO radar system, where electromagnetic vector sensors are employed at each receiver to exploit polarization information of reflected signal. Based on the refined quaternion model, quaternion adaptive detector and its complex counterpart are compared and analyzed. The quaternion adaptive detector outperforms its complex counterpart. The improved performance derives from the better estimate of noise covariance matrix. Further, the accuracy of the estimate is evaluated by Kullback-Leibler (KL) divergence method which can justly compare the quaternion and complex estimates by information entropy. The results are also validated by several numerical simulations.

A fully on-chip linear regulator is presented in this paper to enhance the load transient response in a switch mode power supply (SMPS). Different from the conventional linear regulator, it employs a push-pull output structure, allowing both step-up and step-down transient current compensation. A multi-level self-biased controller is designed to adaptively compensate the instantaneous current difference between the SMPS and the load, achieving ultra-fast and smooth load transient response. An IC prototype is designed and fabricated with a standard 0.35μm digital CMOS process. 40.3ns load transient response time is achieved for a 100-mA load step change. In addition, it is also capable of dynamic voltage scaling (DVS), with a measured tracking speed of about 1 V/μs.

An integrated switching DC-DC power converter is proposed to implement an adaptive wireless powering scheme in implantable biomedical devices. The proposed dual-loop Δ-Σ modulation, error correction and noise shaping techniques work jointly to improve the transient response, noise suppression and line regulation performances. An observation-based double-sampling (OBDS) sensor and a double-sampling half-clock (DSHC) 1-bit A/D converter are also introduced to further increase the signal processing speed without causing any extra cost. The converter was designed with a standard 130-nm CMOS process. The fully transistor-based HSPICE simulation results demonstrate a 2.8/2.2-μs up/down tracking time to a full-range reference voltage step change. A 7.95-mV/V line regulation is achieved, when the supply voltage varies from 1.1 V to 1.5 V. With a 1.2-V supply, the maximum efficiency of 95.5% is measured at 0.9-V output voltage and 90-mW load power, while the efficiency stays above 80% over the entire 170-mW load power range.

This paper presents the behavioral models for operational amplifier (opamp) by using analog hardware description language, Verilog-A. The Opamppsilas behavioral model is built with limited unit-gain bandwidth, slew-rate and nonlinear gain. A hyperbolic tangent model has been used to describe the nonlinearity of the Opamppsilas gain, which provides the error less than 0.26% against the transistor-level implementation. During the simulation of sigma-delta modulator, the switch-capacitor circuits and comparator are implemented in transistor level, simulations are performed at the transistor and behavioral mixed level, thus the error caused by time sequence has been introduced into the simulation results. The comparative results show that the Verilog-A model for Opamp incurs an error of no more than 0.3 dB in the magnitude of harmonics while providing a 15times advantage in the simulation speed with respect to transistor-level implementations.