arXiv subjects

Shu Liu

Publications and source records attributed to Shu Liu.

At least 145 records · Page 8Linked to original sources

arXiv2022

Generalized Few-shot Semantic Segmentation

Training semantic segmentation models requires a large amount of finely annotated data, making it hard to quickly adapt to novel classes not satisfying this condition. Few-Shot Segmentation (FS-Seg) tackles this problem with many constraints. In this paper, we introduce a new benchmark, called Generalized Few-Shot Semantic Segmentation (GFS-Seg), to analyze the generalization ability of simultaneously segmenting the novel categories with very few examples and the base categories with sufficient examples. It is the first study showing that previous representative state-of-the-art FS-Seg methods fall short in GFS-Seg and the performance discrepancy mainly comes from the constrained setting of FS-Seg. To make GFS-Seg tractable, we set up a GFS-Seg baseline that achieves decent performance without structural change on the original model. Then, since context is essential for semantic segmentation, we propose the Context-Aware Prototype Learning (CAPL) that significantly improves performance by 1) leveraging the co-occurrence prior knowledge from support samples, and 2) dynamically enriching contextual information to the classifier, conditioned on the content of each query image. Both two contributions are experimentally shown to have substantial practical merit. Extensive experiments on Pascal-VOC and COCO manifest the effectiveness of CAPL, and CAPL generalizes well to FS-Seg by achieving competitive performance. Code is available at https://github.com/dvlab-research/GFS-Seg.

cs.CV↗

arXiv2022

ResLT: Residual Learning for Long-tailed Recognition

Deep learning algorithms face great challenges with long-tailed data distribution which, however, is quite a common case in real-world scenarios. Previous methods tackle the problem from either the aspect of input space (re-sampling classes with different frequencies) or loss space (re-weighting classes with different weights), suffering from heavy over-fitting to tail classes or hard optimization during training. To alleviate these issues, we propose a more fundamental perspective for long-tailed recognition, i.e., from the aspect of parameter space, and aims to preserve specific capacity for classes with low frequencies. From this perspective, the trivial solution utilizes different branches for the head, medium, and tail classes respectively, and then sums their outputs as the final results is not feasible. Instead, we design the effective residual fusion mechanism -- with one main branch optimized to recognize images from all classes, another two residual branches are gradually fused and optimized to enhance images from medium+tail classes and tail classes respectively. Then the branches are aggregated into final results by additive shortcuts. We test our method on several benchmarks, i.e., long-tailed version of CIFAR-10, CIFAR-100, Places, ImageNet, and iNaturalist 2018. Experimental results manifest the effectiveness of our method. Our code is available at https://github.com/jiequancui/ResLT.

cs.CV↗

arXiv2022

Deuterated ammonia in Galactic massive star-forming regions

We present sensitive observations of NH2D at 110.153599 GHz toward 50 Galactic massive star-forming regions with IRAM 30-m telescope. The NH2D transition is detected toward 36 objects, yielding a detection rate of 72%. Column densities of NH2D, HC3N and C18O for each source are derived by assuming local thermal equilibrium conditions with a fixed excitation temperature. The deuterium ratio of NH$_3$, defined as the abundance ratio of NH2D to NH3, for 19 sources is also obtained with the information of NH3 from the literature. The range of deuterium fractionation bends to be large in the late-stage star-forming regions in this work, with the value from 0.043 to 0.0006. The highest deuterium ratio of NH3 is 0.043 in G081.75+00.78 (DR21). We also find that the deuterium ratio of NH3 increases with the Galactocentric distances and decreases with the line width.

astro-ph.GA↗

arXiv2022

Wasserstein Hamiltonian flow with common noise on graph

We study the Wasserstein Hamiltonian flow with a common noise on the density manifold of a finite graph. Under the framework of stochastic variational principle, we first develop the formulation of stochastic Wasserstein Hamiltonian flow and show the local existence of a unique solution. We also establish a sufficient condition for the global existence of the solution. Consequently, we obtain the global well-posedness for the nonlinear Schrödinger equations with common noise on graph. In addition, using Wong-Zakai approximation of common noise, we prove the existence of the minimizer for an optimal control problem with common noise. We show that its minimizer satisfies the stochastic Wasserstein Hamiltonian flow on graph as well.

math.OC↗

arXiv2022

Stratified Transformer for 3D Point Cloud Segmentation

3D point cloud segmentation has made tremendous progress in recent years. Most current methods focus on aggregating local features, but fail to directly model long-range dependencies. In this paper, we propose Stratified Transformer that is able to capture long-range contexts and demonstrates strong generalization ability and high performance. Specifically, we first put forward a novel key sampling strategy. For each query point, we sample nearby points densely and distant points sparsely as its keys in a stratified way, which enables the model to enlarge the effective receptive field and enjoy long-range contexts at a low computational cost. Also, to combat the challenges posed by irregular point arrangements, we propose first-layer point embedding to aggregate local information, which facilitates convergence and boosts performance. Besides, we adopt contextual relative position encoding to adaptively capture position information. Finally, a memory-efficient implementation is introduced to overcome the issue of varying point numbers in each window. Extensive experiments demonstrate the effectiveness and superiority of our method on S3DIS, ScanNetv2 and ShapeNetPart datasets. Code is available at https://github.com/dvlab-research/Stratified-Transformer.

cs.CV↗

arXiv2022

Widespread subsonic turbulence in Ophiuchus North 1

Supersonic motions are common in molecular clouds. (Sub)sonic turbulence is usually detected toward dense cores and filaments. However, it remains unknown whether (sub)sonic motions at larger scales ($\gtrsim$1~pc) can be present in different environments or not. Located at a distance of about 110 pc, Ophiuchus North 1 (Oph N1) is one of the nearest molecular clouds that allows in-depth investigation of its turbulence properties by large-scale mapping observations of single-dish telescopes. We carried out the $^{12}$CO ($J=1-0$) and C$^{18}$O ($J=1-0$) imaging observations toward Oph N1 with the Purple Mountain Observatory 13.7 m telescope. The observations have an angular resolution of $\sim$55\arcsec (i.e., 0.03~pc). Most of the whole C$^{18}$O emitting regions have Mach numbers of $\lesssim$1, demonstrating the large-scale (sub)sonic turbulence across Oph N1. Based on the polarization measurements, we estimate the magnetic field strength of the plane-of-sky component to be $\gtrsim$9~$μ$G. We infer that Oph N1 is globally sub-Alfv{é}nic, and is supported against gravity mainly by the magnetic field. The steep velocity structure function can be caused by the expansion of the Sh~2-27 H{\scriptsize II} region or the dissipative range of incompressible turbulence. Our observations reveal a surprising case of clouds characterised by widespread subsonic turbulence and steep size-linewidth relationship. This cloud is magnetized where ion-neutral friction should play an important role.

astro-ph.GA↗

arXiv2022

SEA: Bridging the Gap Between One- and Two-stage Detector Distillation via SEmantic-aware Alignment

We revisit the one- and two-stage detector distillation tasks and present a simple and efficient semantic-aware framework to fill the gap between them. We address the pixel-level imbalance problem by designing the category anchor to produce a representative pattern for each category and regularize the topological distance between pixels and category anchors to further tighten their semantic bonds. We name our method SEA (SEmantic-aware Alignment) distillation given the nature of abstracting dense fine-grained information by semantic reliance to well facilitate distillation efficacy. SEA is well adapted to either detection pipeline and achieves new state-of-the-art results on the challenging COCO object detection task on both one- and two-stage detectors. Its superior performance on instance segmentation further manifests the generalization ability. Both 2x-distilled RetinaNet and FCOS with ResNet50-FPN outperform their corresponding 3x ResNet101-FPN teacher, arriving 40.64 and 43.06 AP, respectively. Code will be made publicly available.

cs.CV↗

arXiv2021

Stochastic Wasserstein Hamiltonian Flows

In this paper, we study the stochastic Hamiltonian flow in Wasserstein manifold, the probability density space equipped with $L^2$-Wasserstein metric tensor, via the Wong--Zakai approximation. We begin our investigation by showing that the stochastic Euler-Lagrange equation, regardless it is deduced from either variational principle or particle dynamics, can be interpreted as the stochastic kinetic Hamiltonian flows in Wasserstein manifold. We further propose a novel variational formulation to derive more general stochastic Wassersetin Hamiltonian flows, and demonstrate that this new formulation is applicable to various systems including the stochastic Schrödinger equation, Schrödinger equation with random dispersion, and Schrödinger bridge problem with common noise.

math.PR↗

arXiv2021

Bounds and Constructions for Insertion and Deletion Codes

The present paper mainly studies limits and constructions of insertion and deletion (insdel for short) codes. The paper can be divided into two parts. The first part focuses on various bounds, while the second part concentrates on constructions of insdel codes. Although the insdel-metric Singleton bound has been derived before, it is still unknown if there are any nontrivial codes achieving this bound. Our first result shows that any nontrivial insdel codes do not achieve the insdel-metric Singleton bound. The second bound shows that every $[n,k]$ Reed-Solomon code has insdel distance upper bounded by $2n-4k+4$ and it is known in literature that an $[n,k]$ Reed-Solomon code can have insdel distance $2n-4k+4$ as long as the field size is sufficiently large. The third bound shows a trade-off between insdel distance and code alphabet size for codes achieving the Hamming-metric Singleton bound. In the second part of the paper, we first provide a non-explicit construction of nonlinear codes that can approach the insdel-metric Singleton bound arbitrarily when the code alphabet size is sufficiently large. The second construction gives two-dimensional Reed-Solomon codes of length $n$ and insdel distance $2n-4$ with field size $q=O(n^5)$.

cs.IT↗

arXiv2021

The Dependence of the Type Ia Supernova Host Bias on Observation or Fitting Technique

More luminous Type Ia supernovae (SNe Ia) prefer less massive hosts and regions of higher star formation. This correlation is inverted during width-color-luminosity light curve standardization resulting in step-like biases of distance measurements with respect to host properties. Using the PISCO supernova host sample and SDSS, GALEX, and 2MASS photometry, we compare host stellar mass and specific star formation rate (sSFR) from different observation methods, including local vs. global, and fitting techniques to measure their impact on the host step biases. Mass step measurements for all our mass samples are consistent within a 1$σ$ significance from -0.03$\pm$0.02 mag to -0.04$\pm$0.02 mag. Including or excluding UV information had no effect on measured mass step size or location. Specific SFR (sSFR) step sizes are more significant than mass step measurements and varied from $0.05\pm0.03$ mag (H$α$) and $0.06\pm0.02$ mag (UV) for a 51 host sample. The sSFR step location is influenced by mass sample used to normalize star formation and by sSFR tracer choice. The step size is reduced to 0.04$\pm$0.03 mag when using all available 73 hosts with H$α$ measurements. This 73 PISCO host subsample overall lacked a clear step signal, but here we are searching for whether different choices of mass or sSFR estimation can create a step signal. We find no evidence that different observation or fitting techniques choice can create a distance measurement step in either mass or sSFR.

astro-ph.GA↗

arXiv2021

Maximally recoverable local reconstruction codes from subspace direct sum systems

Maximally recoverable local reconstruction codes (MR LRCs for short) have received great attention in the last few years. Various constructions have been proposed in literatures. The main focus of this topic is to construct MR LRCs over small fields. An $(N=nr,r,h,\Gd)$-MR LRC is a linear code over finite field $\F_\ell$ of length $N$, whose codeword symbols are partitioned into $n$ local groups each of size $r$. Each local group can repair any $\Gd$ erasure errors and there are further $h$ global parity checks to provide fault tolerance from more global erasure patterns. MR LRCs deployed in practice have a small number of global parities such as $h=O(1)$. In this parameter setting, all previous constructions require the field size $\ell =Ω_h (N^{h-1-o(1)})$. It remains challenging to improve this bound. In this paper, via subspace direct sum systems, we present a construction of MR LRC with the field size $\ell= O(N^{h-2+\frac1{h-1}-o(1)})$. In particular, for the most interesting cases where $h=2,3$, we improve previous constructions by either reducing field size or removing constraints. In addition, we also offer some constructions of MR LRCs for larger global parity $h$ that have field size incomparable with known upper bounds. The main techniques used in this paper is through subspace direct sum systems that we introduce. Interestingly, subspace direct sum systems are actually equivalent to $\F_q$-linear codes over extension fields. Based on various constructions of subspace direct sum systems, we are able to construct several classes of MR LRCs.

cs.IT↗

arXiv2021

Guided Point Contrastive Learning for Semi-supervised Point Cloud Semantic Segmentation

Rapid progress in 3D semantic segmentation is inseparable from the advances of deep network models, which highly rely on large-scale annotated data for training. To address the high cost and challenges of 3D point-level labeling, we present a method for semi-supervised point cloud semantic segmentation to adopt unlabeled point clouds in training to boost the model performance. Inspired by the recent contrastive loss in self-supervised tasks, we propose the guided point contrastive loss to enhance the feature representation and model generalization ability in semi-supervised setting. Semantic predictions on unlabeled point clouds serve as pseudo-label guidance in our loss to avoid negative pairs in the same category. Also, we design the confidence guidance to ensure high-quality feature learning. Besides, a category-balanced sampling strategy is proposed to collect positive and negative samples to mitigate the class imbalance problem. Extensive experiments on three datasets (ScanNet V2, S3DIS, and SemanticKITTI) show the effectiveness of our semi-supervised method to improve the prediction quality with unlabeled data.

cs.CV↗

arXiv2021

Are Type Ia Supernovae in Restframe $H$ Brighter in More Massive Galaxies?

We analyze 143 Type Ia supernovae (SNeIa) observed in $H$ band (1.6-1.8 $μ$m) and find SNeIa are intrinsically brighter in $H$-band with increasing host galaxy stellar mass. We find SNeIa in galaxies more massive than $10^{10.43} M_{\odot}$ are $0.13 \pm 0.04$ mag brighter in $H$ than SNeIa in less massive galaxies. The same set of SNeIa observed at optical wavelengths, after width-color-luminosity corrections, exhibit a $0.10 \pm 0.03$ mag offset in the Hubble residuals. We observe an outlier population ($|ΔH_{\rm max}| > 0.5$ mag) in the $H$ band and show that removing the outlier population moves the mass threshold to $10^{10.65} M_{\odot}$ and reduces the step in $H$ band to $0.08 \pm 0.04$ mag, but the equivalent optical mass step is increased to $0.13 \pm 0.04$ mag. We conclude the outliers do not drive the brightness--host-mass correlation. Less massive galaxies preferentially host more higher-stretch SNeIa, which are intrinsically brighter and bluer. It is only after correction for width-luminosity and color-luminosity relationships that SNeIa have brighter optical Hubble residuals in more massive galaxies. Thus finding SNeIa are intrinsically brighter in $H$ in more massive galaxies is an opposite correlation to the intrinsic (pre-width-luminosity correction) optical brightness. If dust and the treatment of intrinsic color variation were the main driver of the host galaxy mass correlation, we would not expect a correlation of brighter $H$-band SNeIa in more massive galaxies.

astro-ph.GA↗

arXiv2021

Video Instance Segmentation with a Propose-Reduce Paradigm

Video instance segmentation (VIS) aims to segment and associate all instances of predefined classes for each frame in videos. Prior methods usually obtain segmentation for a frame or clip first, and merge the incomplete results by tracking or matching. These methods may cause error accumulation in the merging step. Contrarily, we propose a new paradigm -- Propose-Reduce, to generate complete sequences for input videos by a single step. We further build a sequence propagation head on the existing image-level instance segmentation network for long-term propagation. To ensure robustness and high recall of our proposed framework, multiple sequences are proposed where redundant sequences of the same instance are reduced. We achieve state-of-the-art performance on two representative benchmark datasets -- we obtain 47.6% in terms of AP on YouTube-VIS validation set and 70.4% for J&F on DAVIS-UVOS validation set. Code is available at https://github.com/dvlab-research/ProposeReduce.

cs.CV↗

arXiv2021

Deep Structured Instance Graph for Distilling Object Detectors

Effectively structuring deep knowledge plays a pivotal role in transfer from teacher to student, especially in semantic vision tasks. In this paper, we present a simple knowledge structure to exploit and encode information inside the detection system to facilitate detector knowledge distillation. Specifically, aiming at solving the feature imbalance problem while further excavating the missing relation inside semantic instances, we design a graph whose nodes correspond to instance proposal-level features and edges represent the relation between nodes. To further refine this graph, we design an adaptive background loss weight to reduce node noise and background samples mining to prune trivial edges. We transfer the entire graph as encoded knowledge representation from teacher to student, capturing local and global information simultaneously. We achieve new state-of-the-art results on the challenging COCO object detection task with diverse student-teacher pairs on both one- and two-stage detectors. We also experiment with instance segmentation to demonstrate robustness of our method. It is notable that distilled Faster R-CNN with ResNet18-FPN and ResNet50-FPN yields 38.68 and 41.82 Box AP respectively on the COCO benchmark, Faster R-CNN with ResNet101-FPN significantly achieves 43.38 AP, which outperforms ResNet152-FPN teacher about 0.7 AP. Code: https://github.com/dvlab-research/Dsig.

cs.CV↗

arXiv2021

Exploring and Improving Mobile Level Vision Transformers

We study the vision transformer structure in the mobile level in this paper, and find a dramatic performance drop. We analyze the reason behind this phenomenon, and propose a novel irregular patch embedding module and adaptive patch fusion module to improve the performance. We conjecture that the vision transformer blocks (which consist of multi-head attention and feed-forward network) are more suitable to handle high-level information than low-level features. The irregular patch embedding module extracts patches that contain rich high-level information with different receptive fields. The transformer blocks can obtain the most useful information from these irregular patches. Then the processed patches pass the adaptive patch merging module to get the final features for the classifier. With our proposed improvements, the traditional uniform vision transformer structure can achieve state-of-the-art results in mobile level. We improve the DeiT baseline by more than 9\% under the mobile-level settings and surpass other transformer architectures like Swin and CoaT by a large margin.

cs.CV↗

arXiv2021

Parametric Contrastive Learning

In this paper, we propose Parametric Contrastive Learning (PaCo) to tackle long-tailed recognition. Based on theoretical analysis, we observe supervised contrastive loss tends to bias on high-frequency classes and thus increases the difficulty of imbalanced learning. We introduce a set of parametric class-wise learnable centers to rebalance from an optimization perspective. Further, we analyze our PaCo loss under a balanced setting. Our analysis demonstrates that PaCo can adaptively enhance the intensity of pushing samples of the same class close as more samples are pulled together with their corresponding centers and benefit hard example learning. Experiments on long-tailed CIFAR, ImageNet, Places, and iNaturalist 2018 manifest the new state-of-the-art for long-tailed recognition. On full ImageNet, models trained with PaCo loss surpass supervised contrastive learning across various ResNet backbones, e.g., our ResNet-200 achieves 81.8% top-1 accuracy. Our code is available at https://github.com/dvlab-research/Parametric-Contrastive-Learning.

cs.CV↗

arXiv2021

Learnable Boundary Guided Adversarial Training

Previous adversarial training raises model robustness under the compromise of accuracy on natural data. In this paper, we reduce natural accuracy degradation. We use the model logits from one clean model to guide learning of another one robust model, taking into consideration that logits from the well trained clean model embed the most discriminative features of natural data, {\it e.g.}, generalizable classifier boundary. Our solution is to constrain logits from the robust model that takes adversarial examples as input and makes it similar to those from the clean model fed with corresponding natural data. It lets the robust model inherit the classifier boundary of the clean model. Moreover, we observe such boundary guidance can not only preserve high natural accuracy but also benefit model robustness, which gives new insights and facilitates progress for the adversarial community. Finally, extensive experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet testify to the effectiveness of our method. We achieve new state-of-the-art robustness on CIFAR-100 without additional real or synthetic data with auto-attack benchmark \footnote{\url{https://github.com/fra31/auto-attack}}. Our code is available at \url{https://github.com/dvlab-research/LBGAT}.

cs.CV↗