arXiv subjects

Shu Liu

Publications and source records attributed to Shu Liu.

At least 163 records · Page 9Linked to original sources

arXiv2021

HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration

Point cloud registration is a fundamental problem in 3D computer vision. Outdoor LiDAR point clouds are typically large-scale and complexly distributed, which makes the registration challenging. In this paper, we propose an efficient hierarchical network named HRegNet for large-scale outdoor LiDAR point cloud registration. Instead of using all points in the point clouds, HRegNet performs registration on hierarchically extracted keypoints and descriptors. The overall framework combines the reliable features in deeper layer and the precise position information in shallower layers to achieve robust and precise registration. We present a correspondence network to generate correct and accurate keypoints correspondences. Moreover, bilateral consensus and neighborhood consensus are introduced for keypoints matching and novel similarity features are designed to incorporate them into the correspondence network, which significantly improves the registration performance. Besides, the whole network is also highly efficient since only a small number of keypoints are used for registration. Extensive experiments are conducted on two large-scale outdoor LiDAR point cloud datasets to demonstrate the high accuracy and efficiency of the proposed HRegNet. The project website is https://ispc-group.github.io/hregnet.

cs.CV↗

arXiv2021

Semi-supervised Semantic Segmentation with Directional Context-aware Consistency

Semantic segmentation has made tremendous progress in recent years. However, satisfying performance highly depends on a large number of pixel-level annotations. Therefore, in this paper, we focus on the semi-supervised segmentation problem where only a small set of labeled data is provided with a much larger collection of totally unlabeled images. Nevertheless, due to the limited annotations, models may overly rely on the contexts available in the training data, which causes poor generalization to the scenes unseen before. A preferred high-level representation should capture the contextual information while not losing self-awareness. Therefore, we propose to maintain the context-aware consistency between features of the same identity but with different contexts, making the representations robust to the varying environments. Moreover, we present the Directional Contrastive Loss (DC Loss) to accomplish the consistency in a pixel-to-pixel manner, only requiring the feature with lower quality to be aligned towards its counterpart. In addition, to avoid the false-negative samples and filter the uncertain positive samples, we put forward two sampling strategies. Extensive experiments show that our simple yet effective method surpasses current state-of-the-art methods by a large margin and also generalizes well with extra image-level annotations.

cs.CV↗

arXiv2021

Learning Stochastic Behaviour from Aggregate Data

Learning nonlinear dynamics from aggregate data is a challenging problem because the full trajectory of each individual is not available, namely, the individual observed at one time may not be observed at the next time point, or the identity of individual is unavailable. This is in sharp contrast to learning dynamics with full trajectory data, on which the majority of existing methods are based. We propose a novel method using the weak form of Fokker Planck Equation (FPE) -- a partial differential equation -- to describe the density evolution of data in a sampled form, which is then combined with Wasserstein generative adversarial network (WGAN) in the training process. In such a sample-based framework we are able to learn the nonlinear dynamics from aggregate data without explicitly solving the partial differential equation (PDE) FPE. We demonstrate our approach in the context of a series of synthetic and real-world data sets.

cs.LG↗

arXiv2021

Learning High Dimensional Wasserstein Geodesics

We propose a new formulation and learning strategy for computing the Wasserstein geodesic between two probability distributions in high dimensions. By applying the method of Lagrange multipliers to the dynamic formulation of the optimal transport (OT) problem, we derive a minimax problem whose saddle point is the Wasserstein geodesic. We then parametrize the functions by deep neural networks and design a sample based bidirectional learning algorithm for training. The trained networks enable sampling from the Wasserstein geodesic. As by-products, the algorithm also computes the Wasserstein distance and OT map between the marginal distributions. We demonstrate the performance of our algorithms through a series of experiments with both synthetic and realistic data.

cs.LG↗

arXiv2021

Approximating the Optimal Transport Plan via Particle-Evolving Method

Optimal transport (OT) provides powerful tools for comparing probability measures in various types. The Wasserstein distance which arises naturally from the idea of OT is widely used in many machine learning applications. Unfortunately, computing the Wasserstein distance between two continuous probability measures always suffers from heavy computational intractability. In this paper, we propose an innovative algorithm that iteratively evolves a particle system to match the optimal transport plan for two given continuous probability measures. The derivation of the algorithm is based on the construction of the gradient flow of an Entropy Transport Problem which could be naturally understood as a classical Wasserstein optimal transport problem with relaxed marginal constraints. The algorithm comes with theoretical analysis and empirical evidence.

math.OC↗

arXiv2021

Convergent Filaments Contracting Towards an Intermediate-mass Prestellar Core

Filamentary structures are closely associated with star-forming cores, but their detailed physical connections are still not clear. We studied the dense gas in the region of OMC-3 MMS-7 in Orion A molecular cloud using the molecular lines observed with the Atacama Large Millimeter/submillimeter Array (ALMA) and the Submillimeter Array (SMA). The ALMA N$_2$H$^+$ (1-0) emission has revealed three dense filaments intersected at the center, coincident with the central core MMS-7, which has a mass of $3.6\,M_\odot$. The filaments and cores are embedded in a parental clump with total mass of $29\,M_\odot$. The N$_2$H$^+$ velocity field exhibits a noticeable increasing trend along the filaments towards the central core MMS-7 with a scale of $v-v_{\rm lsr} \simeq 1.5$ ${\rm km\, s^{-1}}$ over a spatial range of $\sim$20 arcsec ($8\times 10^3$ AU), corresponding to a gradient of $40\,{\rm km\, s^{-1}}\,{\rm pc}^{-1}$. This feature is most likely to indicate an infall motion towards the center. The derived infall rate ($8\times 10^{-5}\,M_\odot$ year$^{-1}$) and timescale ($3.6\times 10^5$ years) are much lower than that in a spherical free-fall collapse and more consistent with the contraction of filament structures. The filaments also exhibit a possible fragmentation, but it does not seem to largely interrupt the gas structure or the infall motion towards the center. MMS-7 thus provides an example of filamentary infall into an individual prestellar core. The filament contraction could be less intense but more steady than the global spherical collapse, and may help generate an intermediate- or even high-mass star.

astro-ph.GA↗

arXiv2021

Distilling Knowledge via Knowledge Review

Knowledge distillation transfers knowledge from the teacher network to the student one, with the goal of greatly improving the performance of the student network. Previous methods mostly focus on proposing feature transformation and loss functions between the same level's features to improve the effectiveness. We differently study the factor of connection path cross levels between teacher and student networks, and reveal its great importance. For the first time in knowledge distillation, cross-stage connection paths are proposed. Our new review mechanism is effective and structurally simple. Our finally designed nested and compact framework requires negligible computation overhead, and outperforms other methods on a variety of tasks. We apply our method to classification, object detection, and instance segmentation tasks. All of them witness significant student network performance improvement. Code is available at https://github.com/Jia-Research-Lab/ReviewKD

cs.CV↗

arXiv2021

HCN (1-0) opacity of outflowing gas in Arp 220W

Context. We present our findings for the HCN/H13CN 1-0 line ratio in the molecular outflow of Arp 220 west based on high-resolution ALMA data. Aims. Molecular gas masses in the outflowing gas of galaxies driven by active galactic nuclei (AGNs) or starbursts are important parameters for understanding the feedback of these latter two phenomena and star-formation quenching. The conversion factor of line luminosities to masses is related to the optical depth of the molecular lines. Methods. Using H13CN 1-0, the isotopic line of HCN 1-0, to obtain the line ratio of HCN/H13CN 1-0 is an ideal way to derive the optical depth of HCN 1-0 in outflowing gas. Results. With the nondetection of H13CN 1-0 in the outflowing gas, a 3σ lower limit of HCN/H13CN 1-0 line ratio is obtained, which is at least three times higher than that found in the whole of the whole system of Arp 220. The high HCN/H13CN 1-0 line ratio indicates low opacity of HCN 1-0 in the outflowing gas, even though the upper limit of HCN 1-0 opacity obtained here is still not good enough to draw any robust conclusions if HCN 1-0 is optically thin. A lower conversion factor of HCN 1-0 luminosity to dense gas mass in the outflowing gas should be used than that used for the host galaxy of Arp 220.

astro-ph.GA↗

arXiv2021

Jigsaw Clustering for Unsupervised Visual Representation Learning

Unsupervised representation learning with contrastive learning achieved great success. This line of methods duplicate each training batch to construct contrastive pairs, making each training batch and its augmented version forwarded simultaneously and leading to additional computation. We propose a new jigsaw clustering pretext task in this paper, which only needs to forward each training batch itself, and reduces the training cost. Our method makes use of information from both intra- and inter-images, and outperforms previous single-batch based ones by a large margin. It is even comparable to the contrastive learning methods when only half of training batches are used. Our method indicates that multiple batches during training are not necessary, and opens the door for future research of single-batch unsupervised methods. Our models trained on ImageNet datasets achieve state-of-the-art results with linear classification, outperforming previous single-batch methods by 2.6%. Models transferred to COCO datasets outperform MoCo v2 by 0.4% with only half of the training batches. Our pretrained models outperform supervised ImageNet pretrained models on CIFAR-10 and CIFAR-100 datasets by 0.9% and 4.1% respectively. Code is available at https://github.com/Jia-Research-Lab/JigsawClustering

cs.CV↗

arXiv2021

Improving Calibration for Long-Tailed Recognition

Deep neural networks may perform poorly when training datasets are heavily class-imbalanced. Recently, two-stage methods decouple representation learning and classifier learning to improve performance. But there is still the vital issue of miscalibration. To address it, we design two methods to improve calibration and performance in such scenarios. Motivated by the fact that predicted probability distributions of classes are highly related to the numbers of class instances, we propose label-aware smoothing to deal with different degrees of over-confidence for classes and improve classifier learning. For dataset bias between these two stages due to different samplers, we further propose shifted batch normalization in the decoupling framework. Our proposed methods set new records on multiple popular long-tailed recognition benchmark datasets, including CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist 2018. Code will be available at https://github.com/Jia-Research-Lab/MiSLAS.

cs.CV↗

arXiv2021

Dense gas in local galaxies revealed by multiple tracers

We present 3 mm and 2 mm band simultaneously spectroscopic observations of HCN 1-0, HCO$^{+}$ 1-0, HNC 1-0, and CS 3-2 with the IRAM 30 meter telescope, toward a sample of 70 sources as nearby galaxies with infrared luminosities ranging from several 10$^{5}L_{\odot}$ to more than 10$^{12}L_{\odot}$. After combining HCN 1-0, HCO$^{+}$ 1-0 and HNC 1-0 data from literature with our detections, relations between luminosities of dense gas tracers (HCN 1-0, HCO$^{+}$ 1-0 and HNC 1-0) and infrared luminosities are derived, with tight linear correlations for all tracers. Luminosities of CS 3-2 with only our observations also show tight linear correlation with infrared luminosities. No systematic difference is found for tracing dense molecular gas among these tracers. Star formation efficiencies for dense gas with different tracers also do not show any trend along different infrared luminosities. Our study also shows that HCN/HCO$^{+}$ line ratio might not be a good indicator to diagnose obscured AGN in galaxies.

astro-ph.GA↗

arXiv2020

On List Decoding of Insertion and Deletion Errors

Insdel errors occur in communication systems caused by the loss of positional information of the message. Since the work by Guruswami and Wang, there have been some further investigations on the list decoding of insertion codes, deletion codes and insdel codes. However, unlike classical Hamming metric or even rank-metric, there are still many unsolved problems on list decoding of insdel codes. The contributions of this paper mainly consist of two parts. Firstly, we analyze the list decodability of random insdel codes. We show that list decoding of random insdel codes surpasses the Singleton bound when there are more insertion errors than deletion errors and the alphabet size is sufficiently large. Furthermore, our results reveal the existence of an insdel code that can be list decoded against insdel errors beyond its minimum insdel distance while still having polynomial list size. This provides a more complete picture on the list decodability of insdel codes when both insertion and deletion errors happen. Secondly, we construct a family of explicit insdel codes with efficient list decoding algorithm. As a result, we derive a Zyablov-type bound for insdel errors. Recently, after our results appeared, Guruswami et al. provided a complete solution for another open problem on list decoding of insdel codes. In contrast to the problems we considered, they provided a region containing all possible insertion and deletion errors that are still list decodable by some q-ary insdel codes of non-zero rate. More specifically, for a fixed number of insertion and deletion errors, while our paper focuses on maximizing the rate of a code that is list decodable against that amount of insertion and deletion errors, Guruswami et al. focuses on finding out the existence of a code with asymptotically non-zero rate which is list decodable against this amount of insertion and deletion errors.

cs.IT↗

arXiv2020

CO observations toward HI-rich Ultra Diffuse Galaxies

We present CO observations toward a sample of six HI-rich Ultra-diffuse galaxies (UDGs) as well as one UDG (VLSB-A) in the Virgo Cluster with the IRAM 30-m telescope. CO 1-0 is marginally detected at 4sigma level in AGC122966, as the first detection of CO emission in UDGs. We estimate upper limits of molecular mass in other galaxies from the non-detection of CO lines. These upper limits and the marginal CO detection in AGC122966 indicate low mass ratios between molecular and atomic gas masses. With the star formation efficiency derived from the molecular gas, we suggest that the inefficiency of star formation in such HI-rich UDGs is likely caused by the low efficiency in converting molecules from atomic gas, instead of low efficiency in forming stars from molecular gas.

astro-ph.GA↗

arXiv2020

Dive Deeper Into Box for Object Detection

Anchor free methods have defined the new frontier in state-of-the-art object detection researches where accurate bounding box estimation is the key to the success of these methods. However, even the bounding box has the highest confidence score, it is still far from perfect at localization. To this end, we propose a box reorganization method(DDBNet), which can dive deeper into the box for more accurate localization. At the first step, drifted boxes are filtered out because the contents in these boxes are inconsistent with target semantics. Next, the selected boxes are broken into boundaries, and the well-aligned boundaries are searched and grouped into a sort of optimal boxes toward tightening instances more precisely. Experimental results show that our method is effective which leads to state-of-the-art performance for object detection.

cs.CV↗

arXiv2020

Filament Intersections and Cold Dense Cores in Orion A North

We studied the filament structures and dense cores in OMC-2,3 region in Orion A North molecular cloud using the high-resolution N2H+ (1-0) spectral cube observed with the Atacama Large Millimeter/Submillimeter Array (ALMA). The filament network over a total length of 2 pc is found to contain 170 intersections and 128 candidate dense cores. The dense cores are all displaced from the infrared point sources (possible young stars), and the major fraction of cores (103) are located around the intersections. Towards the intersections, there is also an increasing trend for the total column density Ntot as well as the the power-law index of the column-density Probability Distribution Function (N-PDF), suggesting that the intersections would in general have more significant gas assembly than the other part of the filament paths. The virial analysis shows that the dense cores mostly have virial mass ratio of alpha_vir=M_vir/M_gas<1.0, suggesting them to be bounded by the self gravity. In the mean time, only about 23 percent of the cores have critical mass ratio of alpha_crit=M_crit/M_gas<1.0, suggesting them to be unstable against core collapse. Combining these results, it shows that the major fraction of the cold starless and possible prestellar cores in OMC-2,3 are being assembled around the intersections, and currently in a gravitationally bound state. But more extensive core collapse and star formation may still require continuous core-mass growth or other perturbatio

astro-ph.GA↗

arXiv2020

Parametric Fokker-Planck equation

We derive the Fokker-Planck equation on the parametric space. It is the Wasserstein gradient flow of relative entropy on the statistical manifold. We pull back the PDE to a finite dimensional ODE on parameter space. Some analytical example and numerical examples are presented.

math.OC↗

arXiv2020

Jointly Modeling Aspect and Sentiment with Dynamic Heterogeneous Graph Neural Networks

Target-Based Sentiment Analysis aims to detect the opinion aspects (aspect extraction) and the sentiment polarities (sentiment detection) towards them. Both the previous pipeline and integrated methods fail to precisely model the innate connection between these two objectives. In this paper, we propose a novel dynamic heterogeneous graph to jointly model the two objectives in an explicit way. Both the ordinary words and sentiment labels are treated as nodes in the heterogeneous graph, so that the aspect words can interact with the sentiment information. The graph is initialized with multiple types of dependencies, and dynamically modified during real-time prediction. Experiments on the benchmark datasets show that our model outperforms the state-of-the-art models. Further analysis demonstrates that our model obtains significant performance gain on the challenging instances under multiple-opinion aspects and no-opinion aspect situations.

cs.CL↗

arXiv2020

DSGN: Deep Stereo Geometry Network for 3D Object Detection

Most state-of-the-art 3D object detectors heavily rely on LiDAR sensors because there is a large performance gap between image-based and LiDAR-based methods. It is caused by the way to form representation for the prediction in 3D scenarios. Our method, called Deep Stereo Geometry Network (DSGN), significantly reduces this gap by detecting 3D objects on a differentiable volumetric representation -- 3D geometric volume, which effectively encodes 3D geometric structure for 3D regular space. With this representation, we learn depth information and semantic cues simultaneously. For the first time, we provide a simple and effective one-stage stereo-based 3D detection pipeline that jointly estimates the depth and detects 3D objects in an end-to-end learning manner. Our approach outperforms previous stereo-based 3D detectors (about 10 higher in terms of AP) and even achieves comparable performance with several LiDAR-based methods on the KITTI 3D object detection leaderboard. Our code is publicly available at https://github.com/chenyilun95/DSGN.

cs.CV↗