arXiv subjects

Jian Sun

Publications and source records attributed to Jian Sun.

At least 433 records · Page 24Linked to original sources

arXiv2018

MegDet: A Large Mini-Batch Object Detector

The improvements in recent CNN-based object detection works, from R-CNN [11], Fast/Faster R-CNN [10, 31] to recent Mask R-CNN [14] and RetinaNet [24], mainly come from new network, new framework, or novel loss design. But mini-batch size, a key factor in the training, has not been well studied. In this paper, we propose a Large MiniBatch Object Detector (MegDet) to enable the training with much larger mini-batch size than before (e.g. from 16 to 256), so that we can effectively utilize multiple GPUs (up to 128 in our experiments) to significantly shorten the training time. Technically, we suggest a learning rate policy and Cross-GPU Batch Normalization, which together allow us to successfully train a large mini-batch detector in much less time (e.g., from 33 hours to 4 hours), and achieve even better accuracy. The MegDet is the backbone of our submission (mmAP 52.5%) to COCO 2017 Challenge, where we won the 1st place of Detection task.

cs.CV↗

arXiv2018

ExFuse: Enhancing Feature Fusion for Semantic Segmentation

Modern semantic segmentation frameworks usually combine low-level and high-level features from pre-trained backbone convolutional models to boost performance. In this paper, we first point out that a simple fusion of low-level and high-level features could be less effective because of the gap in semantic levels and spatial resolution. We find that introducing semantic information into low-level features and high-resolution details into high-level features is more effective for the later fusion. Based on this observation, we propose a new framework, named ExFuse, to bridge the gap between low-level and high-level features thus significantly improve the segmentation quality by 4.0\% in total. Furthermore, we evaluate our approach on the challenging PASCAL VOC 2012 segmentation benchmark and achieve 87.9\% mean IoU, which outperforms the previous state-of-the-art results.

cs.CV↗

arXiv2018

Cascaded Pyramid Network for Multi-Person Pose Estimation

The topic of multi-person pose estimation has been largely improved recently, especially with the development of convolutional neural network. However, there still exist a lot of challenging cases, such as occluded keypoints, invisible keypoints and complex background, which cannot be well addressed. In this paper, we present a novel network structure called Cascaded Pyramid Network (CPN) which targets to relieve the problem from these "hard" keypoints. More specifically, our algorithm includes two stages: GlobalNet and RefineNet. GlobalNet is a feature pyramid network which can successfully localize the "simple" keypoints like eyes and hands but may fail to precisely recognize the occluded or invisible keypoints. Our RefineNet tries explicitly handling the "hard" keypoints by integrating all levels of feature representations from the GlobalNet together with an online hard keypoint mining loss. In general, to address the multi-person pose estimation problem, a top-down pipeline is adopted to first generate a set of human bounding boxes based on a detector, followed by our CPN for keypoint localization in each human bounding box. Based on the proposed algorithm, we achieve state-of-art results on the COCO keypoint benchmark, with average precision at 73.0 on the COCO test-dev dataset and 72.1 on the COCO test-challenge dataset, which is a 19% relative improvement compared with 60.5 from the COCO 2016 keypoint challenge.Code (https://github.com/chenyilun95/tf-cpn.git) and the detection results are publicly available for further research.

cs.CV↗

arXiv2018

Unveiling the Charge Density Wave Inhomogeneity and Pseudogap State in 1T-TiSe2

By using scanning tunneling microscopy (STM) / spectroscopy (STS), we systematically characterize the electronic structure of lightly doped 1T-TiSe2, and demonstrate the existence of the electronic inhomogeneity and the pseudogap state. It is found that the intercalation induced lattice distortion impacts the local band structure and reduce the size of the charge density wave (CDW) gap with the persisted 2x2 spatial modulation. On the other hand, the delocalized doping electrons promote the formation of pseudogap. Domination by either of the two effects results in the separation of two characteristic regions in real space, exhibiting rather different electronic structures. Further doping electrons to the surface confirms that the pseudogap may be the precursor for the superconducting gap. This study suggests that the competition of local lattice distortion and the delocalized doping effect contribute to the complicated relationship between charge density wave and superconductivity for intercalated 1T-TiSe2.

cond-mat.str-el↗

arXiv2018

Repulsion Loss: Detecting Pedestrians in a Crowd

Detecting individual pedestrians in a crowd remains a challenging problem since the pedestrians often gather together and occlude each other in real-world scenarios. In this paper, we first explore how a state-of-the-art pedestrian detector is harmed by crowd occlusion via experimentation, providing insights into the crowd occlusion problem. Then, we propose a novel bounding box regression loss specifically designed for crowd scenes, termed repulsion loss. This loss is driven by two motivations: the attraction by target, and the repulsion by other surrounding objects. The repulsion term prevents the proposal from shifting to surrounding objects thus leading to more crowd-robust localization. Our detector trained by repulsion loss outperforms all the state-of-the-art methods with a significant improvement in occlusion cases.

cs.CV↗

arXiv2018

A novel superhard tungsten nitride predicted by machine-learning accelerated crystal structure searching

Transition metal nitrides have been suggested to have both high hardness and good thermal stability with large potential application value, but so far stable superhard transition metal nitrides have not been synthesized. Here, with our newly developed machine-learning accelerated crystal structure searching method, we designed a superhard tungsten nitride, h-WN6, which can be synthesized at pressure around 65 GPa and quenchable to ambient pressure. This h-WN6 is constructed with single-bonded N6 rings and presents ionic-like features, which can be formulated as W2.4+N62.4-. It has a band gap of 1.6 eV at 0 GPa and exhibits an abnormal gap broadening behavior under pressure. Excitingly, this h-WN6 is found to be the hardest among transition metal nitrides known so far (Vickers hardness around 57 GPa) and also has a very high melting temperature (around 1900 K). These predictions support the designing rules and may stimulate future experiments to synthesize superhard material.

cond-mat.mtrl-sci↗

arXiv2018

AlignedReID: Surpassing Human-Level Performance in Person Re-Identification

In this paper, we propose a novel method called AlignedReID that extracts a global feature which is jointly learned with local features. Global feature learning benefits greatly from local feature learning, which performs an alignment/matching by calculating the shortest path between two sets of local features, without requiring extra supervision. After the joint learning, we only keep the global feature to compute the similarities between images. Our method achieves rank-1 accuracy of 94.4% on Market1501 and 97.8% on CUHK03, outperforming state-of-the-art methods by a large margin. We also evaluate human-level performance and demonstrate that our method is the first to surpass human-level performance on Market1501 and CUHK03, two widely used Person ReID datasets.

cs.CV↗

arXiv2017

Superhard and Superconducting B6C

Crystal structure searching and ab initio calculations have been used here to explore low-energy structures of boron carbides under high pressure. Under pressures of 85-110 GPa, a metastable B6C with R-3m symmetry is found to be energetically more stable than the mixture of previous B4C and elemental boron. This B6C is a rhombohedral structure and contains mooncake-like B24 clusters with stuffing of C2 pairs. The mechanical and dynamical stabilities of this structure at ambient pressure are confirmed by its elastic constants and phonon dispersions. The bulk modulus and shear modulus of this newly predicted B6C at ambient pressure reach high values of 291 GPa and 272 GPa, respectively. The Vickers hardness is calculated to be around 48 GPa, and its melting temperature at ambient pressure is estimated to be around 2500 K, indicating that this metastable B6C is a potential superhard material with very good thermal stability. Interestingly, this superhard B6C structure is also found to be superconducting and its superconducting critical temperature (Tc) is evaluated to be around 12.5 K at ambient pressure by electron-phonon coupling.

cond-mat.mtrl-sci↗

arXiv2017

Nonsymmorphic symmetry protected node-line semimetal in the trigonal YH3

Using ab initio calculations based on density-functional theory and effective model analysis, we propose that the trigonal YH3(Space Group: P-3c1) at ambient pressure is a node-line semimetal when spin-orbit coupling (SOC) is ignored. This trigonal YH3 has very clean electronic structure near Fermi level and its nodal lines locate very closely to the Fermi energy, which makes it a perfect system for model analysis. Symmetry analysis shows that the nodal ring in this compound is protected by the glide-plane symmetry, where the band inversion of |Y+,dxz> and |H1-,s> orbits at Gamma point is responsible for the formation of the nodal lines. When SOC is included, the line nodes are prohibited by the glide-plane symmetry, and a small gap (~5 meV) appears, which leads YH3 to be a strong topological insulator with Z2 indices (1,000). Thus the glide-plane symmetry plays an opposite role in the formation of the nodal lines in cases without and with SOC. As the SOC-induced gap is so small that can be neglected, this P-3c1 YH3 may be a good candidate for experimental explorations on the fundamental physics of topological node-line semimetals. We find the surface states of this P-3c1 phase are somehow unique and may be helpful to identify the real ground state of YH3 in the experiment.

cond-mat.mtrl-sci↗

arXiv2017

Silicon clathrates for photovoltaics predicted by a two-step crystal structure search

Silicon in a cubic diamond structure currently plays a significant role in the photovoltaic industry. However, the intrinsic band structures of crystalline silicon restrict its sunlight conversion efficiency. Recently, a clathrate-like Si-24 has been successfully synthesized, which has a quasi-direct bandgap and sheds light on silicon-based photovoltaics. Here, we proposed a two-step crystal structure search method based on first-principles calculations and explored silicon clathrate structures extensively. First, the guest-host compounds were searched at high pressure, and then, the porous guest-free silicon clathrates were obtained by removing the guest atoms. Using potassium as the guest atom, we identified four metastable silicon clathrate structures, and some of them have bandgaps close to the optimal range of the Shockley-Queisser limit and have a better absorption rate than the cubic diamond silicon. These silicon clathrates may have promising value in photovoltaic applications.

cond-mat.mtrl-sci↗

arXiv2017

ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices

We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new architecture utilizes two new operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error (absolute 7.8%) than recent MobileNet on ImageNet classification task, under the computation budget of 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves ~13x actual speedup over AlexNet while maintaining comparable accuracy.

cs.CV↗

arXiv2017

Light-Head R-CNN: In Defense of Two-Stage Object Detector

In this paper, we first investigate why typical two-stage methods are not as fast as single-stage, fast detectors like YOLO and SSD. We find that Faster R-CNN and R-FCN perform an intensive computation after or before RoI warping. Faster R-CNN involves two fully connected layers for RoI recognition, while R-FCN produces a large score maps. Thus, the speed of these networks is slow due to the heavy-head design in the architecture. Even if we significantly reduce the base model, the computation cost cannot be largely decreased accordingly. We propose a new two-stage detector, Light-Head R-CNN, to address the shortcoming in current two-stage approaches. In our design, we make the head of network as light as possible, by using a thin feature map and a cheap R-CNN subnet (pooling and single fully-connected layer). Our ResNet-101 based light-head R-CNN outperforms state-of-art object detectors on COCO while keeping time efficiency. More importantly, simply replacing the backbone with a tiny network (e.g, Xception), our Light-Head R-CNN gets 30.7 mmAP at 102 FPS on COCO, significantly outperforming the single-stage, fast detectors like YOLO and SSD on both speed and accuracy. Code will be made publicly available.

cs.CV↗

arXiv2017

Channel Pruning for Accelerating Very Deep Neural Networks

In this paper, we introduce a new channel pruning method to accelerate very deep convolutional neural networks.Given a trained CNN model, we propose an iterative two-step algorithm to effectively prune each layer, by a LASSO regression based channel selection and least square reconstruction. We further generalize this algorithm to multi-layer and multi-branch cases. Our method reduces the accumulated error and enhance the compatibility with various architectures. Our pruned VGG-16 achieves the state-of-the-art results by 5x speed-up along with only 0.3% increase of error. More importantly, our method is able to accelerate modern networks like ResNet, Xception and suffers only 1.4%, 1.0% accuracy loss under 2x speed-up respectively, which is significant. Code has been made publicly available.

cs.CV↗

arXiv2017

High-pressure phases of Weyl semimetals NbP, NbAs, TaP and TaAs

The high pressure phase diagrams of TaAs family (NbP, NbAs, TaP and TaAs) are explored systematically using crystal structure search together with first-principles calculations. Our calculations show that NbAs and TaAs have similar phase diagrams and possess the same structural phase transition sequence of I41md $\rightarrow$ P-6m2 $\rightarrow$ P21/c $\rightarrow$ Pm-3m, with a bit different transition pressures. The phase transition sequence of NbP and TaP are somewhat different from that of NbAs and TaAs. In which new structures emerge, for instance, the Cmcm structure in NbP and the Pmmn structure in TaP. It is found interestingly that in the electronic structure of the high-pressure phase P-6m2-NbAs, there are Weyl points and triple degenerate points coexisted, similar to the high-pressure P-6m2-TaAs.

cond-mat.mtrl-sci↗

arXiv2017

High temperature superconducting phase of HBr under pressure predicted by first-principles calculations

The high pressure phases of HBr are explored with an ab initio crystal structure search. By taking into account the contribution of zero-point energy (ZPE), we find that the P4/nmm phase of HBr is thermodynamically stable in the pressure range from 150 to 200 GPa. The superconducting critical temperature (Tc) of P4/nmm HBr is evaluated to be around 73 K at 170 GPa, which is the highest record so far among binary halogen hydrides. Its Tc can be further raised to around 95K under 170 GPa if half of the bromine atoms in the P4/nmm HBr are substituted by the lighter chlorine atoms. Our study shows that, in addition to lower mass, higher coordination number, shorter bonds, and more highly symmetric environment for the hydrogen atoms are important factors to enhance the superconductivity in hydrides.

cond-mat.supr-con↗

arXiv2017

Strain-induced quantum topological phase transitions in Na3Bi

Strain can be used as an effective tool to tune the crystal structure of materials and hence to modify their electronic structures, including topological properties. Here, taking Na3Bi as a paradigmatic example, we demonstrated with first-principles calculations and k$\cdot$p models that the topological phase transitions can be induced by various types of strains. For instance, the Dirac semimetal phase of ambient Na3Bi can be tuned into a topological insulator (TI) phase by uniaxial strain along the h100i axis. Hydrostatic pressure can let the ambient structure transfer into a new thermodynamically stable phase with Fm-3m symmetry, coming with a perfect parabolic semimetal having a single contact point between the conduction and valence bands, exactly at $Γ$ point on the Fermi level like $α$-Sn. Furthermore, uniaxial strain in the <100> direction can tune the new parabolic semimetal phase into a Dirac semimetal, while shear strains in both the <100> and <111> directions can take the new parabolic semimetal phase into a TI. k$\cdot$p models are constructed to gain more insights into these quantum topological phase transitions. At last, we calculated surface states of Fm-3m Na3Bi without and with strains to verify these topological transitions.

cond-mat.mtrl-sci↗

arXiv2017

Direct Visualization of 2D Topological Insulator in Single-layer 1T'-WTe2

We grow nearly freestanding single-layer 1T'-WTe2 on graphitized 6H-SiC(0001) by using molecular beam epitaxy (MBE), and characterize its electronic structure with scanning tunneling microscopy / spectroscopy (STM/STS). We demonstrate the existence of topological edge states at the periphery of single-layer WTe2 islands. Surprisingly, we also find a band gap in the bulk and the semiconducting behaviors of the single-layer WTe2 at low temperature, which is likely resulted from an incommensurate charge density wave (CDW) transition. The realization of two-dimensional topological insulators (2D TIs) in single-layer transition metal dichalcogenide (TMD) thus provides a promising platform for further exploration of the 2D TIs' physics and related applications.

cond-mat.mtrl-sci↗

arXiv2017

ADMM-Net: A Deep Learning Approach for Compressive Sensing MRI

Compressive sensing (CS) is an effective approach for fast Magnetic Resonance Imaging (MRI). It aims at reconstructing MR images from a small number of under-sampled data in k-space, and accelerating the data acquisition in MRI. To improve the current MRI system in reconstruction accuracy and speed, in this paper, we propose two novel deep architectures, dubbed ADMM-Nets in basic and generalized versions. ADMM-Nets are defined over data flow graphs, which are derived from the iterative procedures in Alternating Direction Method of Multipliers (ADMM) algorithm for optimizing a general CS-based MRI model. They take the sampled k-space data as inputs and output reconstructed MR images. Moreover, we extend our network to cope with complex-valued MR images. In the training phase, all parameters of the nets, e.g., transforms, shrinkage functions, etc., are discriminatively trained end-to-end. In the testing phase, they have computational overhead similar to ADMM algorithm but use optimized parameters learned from the data for CS-based reconstruction task. We investigate different configurations in network structures and conduct extensive experiments on MR image reconstruction under different sampling rates. Due to the combination of the advantages in model-based approach and deep learning approach, the ADMM-Nets achieve state-of-the-art reconstruction accuracies with fast computational speed.

cs.CV↗