arXiv subjects

Jianfeng Wang

Publications and source records attributed to Jianfeng Wang.

At least 145 records · Page 8Linked to original sources

arXiv2020

Hashing-based Non-Maximum Suppression for Crowded Object Detection

In this paper, we propose an algorithm, named hashing-based non-maximum suppression (HNMS) to efficiently suppress the non-maximum boxes for object detection. Non-maximum suppression (NMS) is an essential component to suppress the boxes at closely located locations with similar shapes. The time cost tends to be huge when the number of boxes becomes large, especially for crowded scenes. The basic idea of HNMS is to firstly map each box to a discrete code (hash cell) and then remove the boxes with lower confidences if they are in the same cell. Considering the intersection-over-union (IoU) as the metric, we propose a simple yet effective hashing algorithm, named IoUHash, which guarantees that the boxes within the same cell are close enough by a lower IoU bound. For two-stage detectors, we replace NMS in region proposal network with HNMS, and observe significant speed-up with comparable accuracy. For one-stage detectors, HNMS is used as a pre-filter to speed up the suppression with a large margin. Extensive experiments are conducted on CARPK, SKU-110K, CrowdHuman datasets to demonstrate the efficiency and effectiveness of HNMS. Code is released at \url{https://github.com/microsoft/hnms.git}.

cs.CV↗

arXiv2020

Anchor Box Optimization for Object Detection

In this paper, we propose a general approach to optimize anchor boxes for object detection. Nowadays, anchor boxes are widely adopted in state-of-the-art detection frameworks. However, these frameworks usually pre-define anchor box shapes in heuristic ways and fix the sizes during training. To improve the accuracy and reduce the effort of designing anchor boxes, we propose to dynamically learn the anchor shapes, which allows the anchors to automatically adapt to the data distribution and the network learning capability. The learning approach can be easily implemented with stochastic gradient descent and can be plugged into any anchor box-based detection framework. The extra training cost is almost negligible and it has no impact on the inference time or memory cost. Exhaustive experiments demonstrate that the proposed anchor optimization method consistently achieves significant improvement ($\ge 1\%$ mAP absolute gain) over the baseline methods on several benchmark datasets including Pascal VOC 07+12, MS COCO and Brainwash. Meanwhile, the robustness is also verified towards different anchor initialization methods and the number of anchor shapes, which greatly simplifies the problem of anchor box design.

cs.CV↗

arXiv2019

Facial Motion Prior Networks for Facial Expression Recognition

Deep learning based facial expression recognition (FER) has received a lot of attention in the past few years. Most of the existing deep learning based FER methods do not consider domain knowledge well, which thereby fail to extract representative features. In this work, we propose a novel FER framework, named Facial Motion Prior Networks (FMPN). Particularly, we introduce an addition branch to generate a facial mask so as to focus on facial muscle moving regions. To guide the facial mask learning, we propose to incorporate prior domain knowledge by using the average differences between neutral faces and the corresponding expressive faces as the training guidance. Extensive experiments on three facial expression benchmark datasets demonstrate the effectiveness of the proposed method, compared with the state-of-the-art approaches.

cs.CV↗

arXiv2019

Realization of Lieb Lattice in Covalent-organic Frameworks with Tunable Topology and Magnetism

Lieb lattice, a two-dimensional edge-depleted square lattice, has been predicted to host various exotic electronic properties due to its unusual band structure, i.e., Dirac cone intersected by a flat band (Dirac-flat bands). Until now, although a few artificial Lieb lattices have been discovered in experiments, the realization of a Lieb lattice in a real material is still unachievable. In this article, based on tight-binding modeling and first-principles calculations, we predict that the two covalent organic frameworks (COFs), i.e., sp2C-COF and sp2N-COF, which have been synthesized in the recent experiments, are actually the first two material realizations of organic-ligand-based Lieb lattice. It is found that the lattice distortion can govern the bandwidth of the Dirac-flat bands and in turn determine its electronic instability against spontaneous spin-polarization during carrier doping. The spin-orbit coupling effects could drive these Dirac-flat bands in a distorted Lieb lattice presenting nontrivial topological properties, which depend on the position of Fermi level. Interestingly, as the hole doping concentration increases, the sp2C-COF can experience the phase transitions from a paramagnetic state to a ferromagnetic one and then to a Néel antiferromagnetic one. Our findings not only confirm the first material realization of Lieb lattice in COFs, but also offer a possible way to achieve tunable topology and magnetism in d- (f-) orbital-free organic lattices.

cond-mat.mtrl-sci↗

arXiv2019

Statistical inference for block sparsity of complex signals

Block sparsity is an important parameter in many algorithms to successfully recover block sparse signals under the framework of compressive sensing. However, it is often unknown and needs to be estimated. Recently there emerges a few research work about how to estimate block sparsity of real-valued signals, while there is, to the best of our knowledge, no investigation that has been conducted for complex-valued signals. In this paper, we propose a new method to estimate the block sparsity of complex-valued signal. Its statistical properties are obtained and verified by simulations. In addition, we demonstrate the importance of accurately estimating the block sparsity in signal recovery through a sensitivity analysis.

eess.SP↗

arXiv2019

Enhanced block sparse signal recovery based on $q$-ratio block constrained minimal singular values

In this paper we introduce the $q$-ratio block constrained minimal singular values (BCMSV) as a new measure of measurement matrix in compressive sensing of block sparse/compressive signals and present an algorithm for computing this new measure. Both the mixed $\ell_2/\ell_q$ and the mixed $\ell_2/\ell_1$ norms of the reconstruction errors for stable and robust recovery using block Basis Pursuit (BBP), the block Dantzig selector (BDS) and the group lasso in terms of the $q$-ratio BCMSV are investigated. We establish a sufficient condition based on the $q$-ratio block sparsity for the exact recovery from the noise free BBP and developed a convex-concave procedure to solve the corresponding non-convex problem in the condition. Furthermore, we prove that for sub-Gaussian random matrices, the $q$-ratio BCMSV is bounded away from zero with high probability when the number of measurements is reasonably large. Numerical experiments are implemented to illustrate the theoretical results. In addition, we demonstrate that the $q$-ratio BCMSV based error bounds are tighter than the block restricted isotropic constant based bounds.

eess.SP↗

arXiv2019

Giant Enhancement of Solid Solubility in Monolayer BNC Alloys by Selective Orbital Coupling

Solid solubility (SS) is one of the most important features of alloys, which is usually difficult to be largely tuned in the entire alloy concentrations by external approaches. Some alloys that were supposed to have promising physical properties could turn out to be much less useful because of their poor SS, e.g., the case for monolayer BNC [(BN)1-x(C2)x] alloys. Until now, an effective approach on significantly enhancing SS of (BN)1-x(C2)x in the entire x is still lacking. In this article, a novel mechanism of selective orbital coupling between high energy wrong-bond states and surface states mediated by the specific substrate has been proposed to stabilize the wrong-bonds and in turn significantly enhance the SS of (BN)1-x(C2)x alloys. Surprisingly, we demonstrate that five ordered alloys, exhibiting variable direct quasi-particle bandgaps from 1.35 to 3.99 eV, can spontaneously be formed at different x when (BN)1-x(C2)x is grown on hcp-phase Cr. Interestingly, the optical transitions around the band edges in these ordered alloys, accompanied by largely tunable exciton binding energies of ~1 eV at different x, are significantly strong due to their unique band structures. Importantly, the disordered (BN)1-x(C2)x alloys, exhibiting fully tunable bandgaps from 0 to ~6 eV in the entire x, can be formed on Cr substrate at the miscibility temperature of ~1200 K, which is greatly reduced compared to that of 4500~5600 K in free-standing form or on other substrates. Our discovery not only may resolve the long-standing SS problem of BNC alloys, but also could significantly extend the applications of BNC alloys for various optoelectronic applications.

cond-mat.mtrl-sci↗

arXiv2019

Open problems in the spectral theory of signed graphs

Signed graphs are graphs whose edges get a sign $+1$ or $-1$ (the signature). Signed graphs can be studied by means of graph matrices extended to signed graphs in a natural way. Recently, the spectra of signed graphs have attracted much attention from graph spectra specialists. One motivation is that the spectral theory of signed graphs elegantly generalizes the spectral theories of unsigned graphs. On the other hand, unsigned graphs do not disappear completely, since their role can be taken by the special case of balanced signed graphs. Therefore, spectral problems defined and studied for unsigned graphs can be considered in terms of signed graphs, and sometimes such generalization shows nice properties which cannot be appreciated in terms of (unsigned) graphs. Here, we survey some general results on the adjacency spectra of signed graphs, and we consider some spectral problems which are inspired from the spectral theory of (unsigned) graphs.

math.CO↗

arXiv2019

Learning to Count Objects with Few Exemplar Annotations

In this paper, we study the problem of object counting with incomplete annotations. Based on the observation that in many object counting problems the target objects are normally repeated and highly similar to each other, we are particularly interested in the setting when only a few exemplar annotations are provided. Directly applying object detection with incomplete annotations will result in severe accuracy degradation due to its improper handling of unlabeled object instances. To address the problem, we propose a positiveness-focused object detector (PFOD) to progressively propagate the incomplete labels before applying the general object detection algorithm. The PFOD focuses on the positive samples and ignore the negative instances at most of the learning time. This strategy, though simple, dramatically boosts the object counting accuracy. On the CARPK dataset for parking lot car counting, we improved mAP@0.5 from 4.58% to 72.44% using only 5 training images each with 5 bounding boxes. On the Drink35 dataset for shelf product counting, the mAP@0.5 is improved from 14.16% to 53.73% using 10 training images each with 5 bounding boxes.

cs.CV↗

arXiv2019

The spectrum and automorphism group of the set-inclusion graph

Let $n$, $k$ and $l$ be integers with $1\leq k<l\leq n-1$. The set-inclusion graph $G(n,k,l)$ is the graph whose vertex set consists of all $k$- and $l$-subsets of $[n]=\{1,2,\ldots,n\}$, where two distinct vertices are adjacent if one of them is contained in another. In this paper, we determine the spectrum and automorphism group of $G(n,k,l)$, respectively.

math.CO↗

arXiv2019

High Frequency Residual Learning for Multi-Scale Image Classification

We present a novel high frequency residual learning framework, which leads to a highly efficient multi-scale network (MSNet) architecture for mobile and embedded vision problems. The architecture utilizes two networks: a low resolution network to efficiently approximate low frequency components and a high resolution network to learn high frequency residuals by reusing the upsampled low resolution features. With a classifier calibration module, MSNet can dynamically allocate computation resources during inference to achieve a better speed and accuracy trade-off. We evaluate our methods on the challenging ImageNet-1k dataset and observe consistent improvements over different base networks. On ResNet-18 and MobileNet with alpha=1.0, MSNet gains 1.5% accuracy over both architectures without increasing computations. On the more efficient MobileNet with alpha=0.25, our method gains 3.8% accuracy with the same amount of computations.

cs.CV↗

arXiv2019

Prediction of Stoner-Type Magnetism in Low-Dimensional Electrides

Electrides are special ionic solids with excess cavity-trapped electrons serving as anions. Despite the extensive studies on electrides, the interplay between electrides and magnetism is not well understood due to the lack of stable magnetic electrides, particularly the lack of inorganic magnetic electrides. Here, based on the mechanism of Stoner-type magnetic instability, we propose that in certain electrides the low-dimensionality can facilitate the formation of magnetic ground state because of the enhanced density of states near the Fermi level. To be specific, A5B3 (A = Ca, Sr, Ba; B = As, Sb, Bi) (1D), Sr11Mg2Si10 (0D), Ba7Al10 (0D) and Ba4Al5 (0D) have been identified as stable magnetic electrides with spin-polarization energies of tens to hundreds of meV per formula unit. Especially for Ba5As3, the spin-polarization energy can reach up to 220 meV. Furthermore, we demonstrate that the magnetic moment and spin density mainly derive from the interstitial anionic electrons near the Fermi level. Our work paves a way to the searching of stable magnetic electrides and further exploration of the magnetic properties and related applications in electrides.

cond-mat.mtrl-sci↗

arXiv2019

Anomalous Dirac Plasmons in 1D Topological Electrides

Plasmon opens up the possibility to efficiently couple light and matter at sub-wavelength scales. In general, the plasmon frequency is dependent of carrier density. This dependency, however, renders fundamentally a weak plasmon intensity at low frequency, especially for Dirac plasmon (DP) widely studied in graphene. Here we demonstrate a new type of DP, excited by a Dirac nodal-surface state, which exhibits an anomalously density-independent frequency. Remarkably, we predict realization of anomalous DP (ADP) in 1D topological electrides, such as Ba3CrN3 and Sr3CrN3, by first-principles calculations. The ADPs in both systems have a density-independent frequency and high intensity, and their frequency can be tuned from terahertz to mid-infrared by changing the excitation direction. Furthermore, the intrinsic weak electron-phonon coupling of anionic electrons in electrides affords an added advantage of ultra-low phonon-assisted damping and hence a long lifetime of the ADPs. Our work paves the way to developing novel plasmonic and optoelectronic devices by combining topological physics with electride materials.

cond-mat.mtrl-sci↗

arXiv2019

WIDER Face and Pedestrian Challenge 2018: Methods and Results

This paper presents a review of the 2018 WIDER Challenge on Face and Pedestrian. The challenge focuses on the problem of precise localization of human faces and bodies, and accurate association of identities. It comprises of three tracks: (i) WIDER Face which aims at soliciting new approaches to advance the state-of-the-art in face detection, (ii) WIDER Pedestrian which aims to find effective and efficient approaches to address the problem of pedestrian detection in unconstrained environments, and (iii) WIDER Person Search which presents an exciting challenge of searching persons across 192 movies. In total, 73 teams made valid submissions to the challenge tracks. We summarize the winning solutions for all three tracks. and present discussions on open problems and potential research directions in these topics.

cs.CV↗

arXiv2019

Hierarchically Structured Reinforcement Learning for Topically Coherent Visual Story Generation

We propose a hierarchically structured reinforcement learning approach to address the challenges of planning for generating coherent multi-sentence stories for the visual storytelling task. Within our framework, the task of generating a story given a sequence of images is divided across a two-level hierarchical decoder. The high-level decoder constructs a plan by generating a semantic concept (i.e., topic) for each image in sequence. The low-level decoder generates a sentence for each image using a semantic compositional network, which effectively grounds the sentence generation conditioned on the topic. The two decoders are jointly trained end-to-end using reinforcement learning. We evaluate our model on the visual storytelling (VIST) dataset. Empirical results from both automatic and human evaluations demonstrate that the proposed hierarchically structured reinforced training achieves significantly better performance compared to a strong flat deep reinforcement learning baseline.

cs.CV↗

arXiv2018

Revisit Multinomial Logistic Regression in Deep Learning: Data Dependent Model Initialization for Image Recognition

We study in this paper how to initialize the parameters of multinomial logistic regression (a fully connected layer followed with softmax and cross entropy loss), which is widely used in deep neural network (DNN) models for classification problems. As logistic regression is widely known not having a closed-form solution, it is usually randomly initialized, leading to several deficiencies especially in transfer learning where all the layers except for the last task-specific layer are initialized using a pre-trained model. The deficiencies include slow convergence speed, possibility of stuck in local minimum, and the risk of over-fitting. To address those deficiencies, we first study the properties of logistic regression and propose a closed-form approximate solution named regularized Gaussian classifier (RGC). Then we adopt this approximate solution to initialize the task-specific linear layer and demonstrate superior performance over random initialization in terms of both accuracy and convergence speed on various tasks and datasets. For example, for image classification, our approach can reduce the training time by 10 times and achieve 3.2% gain in accuracy for Flickr-style classification. For object detection, our approach can also be 10 times faster in training for the same accuracy, or 5% better in terms of mAP for VOC 2007 with slightly longer training.

cs.CV↗

arXiv2018

Pseudo Dirac Nodal Sphere: Unusual Electronic Structure and Material Realization

Topological semimetals (TSMs) in which conduction and valence bands cross at zero-dimensional (0D) Dirac nodal points (DNPs) or 1D Dirac nodal lines (DNLs), in 3D momentum space, have recently drawn much attention due to their exotic electronic properties. Here we generalize the TSM state further to a higher-symmetry and higher-dimensional pseudo Dirac nodal sphere (PDNS) state, with the band crossings forming a 2D closed sphere at the Fermi level. The PDNS state is characterized with a spherical backbone consisting of multiple crossing DNLs while band degeneracy in between the DNLs is approximately maintained by weak interactions. It exhibits some unique electronic properties and low-energy excitations, such as collective plasmons different from DNPs and DNLs. Based on crystalline symmetries, we theoretically demonstrate two possible types of PDNS states, and identify all the possible band crossings with pairs of 1D irreducible representations to form the PDNS states in 32 point groups. Importantly, we discover that strained MH3 (M= Y, Ho, Tb, Nd) and Si3N2 are materials candidates to realize these two types of PDNS states, respectively. As a high-symmetry-required state, the PDNS semimetal can be regarded as the "parent phase" for other topological gapped and gapless states.

cond-mat.mes-hall↗

arXiv2018

Investigating the Efficiency of the Beijing Faint Object Spectrograph and Camera (BFOSC) of the Xinglong 2.16-m Reflector

The Beijing Faint Object Spectrograph and Camera (BFOSC) is one of the most important instruments of the 2.16-m telescope of the Xinglong Observatory. Every year there are ~ 20 SCI-papers published based on the observational data of this telescope. In this work, we have systemically measured the total efficiency of the BFOSC of the 2.16-m reflector, based on the observations of two ESO flux standard stars. We have obtained the total efficiencies of the BFOSC instrument of different grisms with various slit widths in almost all ranges, and analysed the factors which effect the efficiency of telescope and spectrograph. For the astronomical observers, the result will be useful for them to select a suitable slit width, depending on their scientific goals and weather conditions during the observation; For the technicians, the result will help them systemically find out the real efficiency of telescope and spectrograph, and further to improve the total efficiency and observing capacity of the telescope technically.

astro-ph.IM↗