Searcharxiv⌕ Search

arXiv subjects

Qiyu Sun

Publications and source records attributed to Qiyu Sun.

At least 55 records · Page 3Linked to original sources
arXiv2022

Learn to Adapt for Monocular Depth Estimation

Monocular depth estimation is one of the fundamental tasks in environmental perception and has achieved tremendous progress in virtue of deep learning. However, the performance of trained models tends to degrade or deteriorate when employed on other new datasets due to the gap between different datasets. Though some methods utilize domain adaptation technologies to jointly train different domains and narrow the gap between them, the trained models cannot generalize to new domains that are not involved in training. To boost the transferability of depth estimation models, we propose an adversarial depth estimation task and train the model in the pipeline of meta-learning. Our proposed adversarial task mitigates the issue of meta-overfitting, since the network is trained in an adversarial manner and aims to extract domain invariant representations. In addition, we propose a constraint to impose upon cross-task depth consistency to compel the depth estimation to be identical in different adversarial tasks, which improves the performance of our method and smoothens the training process. Experiments demonstrate that our method adapts well to new datasets after few training steps during the test procedure.

cs.CV↗
arXiv2022

Aggressive Quadrotor Flight Using Curiosity-Driven Reinforcement Learning

The ability to perform aggressive movements, which are called aggressive flights, is important for quadrotors during navigation. However, aggressive quadrotor flights are still a great challenge to practical applications. The existing solutions to aggressive flights heavily rely on a predefined trajectory, which is a time-consuming preprocessing step. To avoid such path planning, we propose a curiosity-driven reinforcement learning method for aggressive flight missions and a similarity-based curiosity module is introduced to speed up the training procedure. A branch structure exploration (BSE) strategy is also applied to guarantee the robustness of the policy and to ensure the policy trained in simulations can be performed in real-world experiments directly. The experimental results in simulations demonstrate that our reinforcement learning algorithm performs well in aggressive flight tasks, speeds up the convergence process and improves the robustness of the policy. Besides, our algorithm shows a satisfactory simulated to real transferability and performs well in real-world experiments.

cs.RO↗
arXiv2021

A Divide-and-Conquer Algorithm for Distributed Optimization on Networks

In this paper, we consider networks with topologies described by some connected undirected graph ${\mathcal{G}}=(V, E)$ and with some agents (fusion centers) equipped with processing power and local peer-to-peer communication, and optimization problem $\min_{\boldsymbol x}\big\{F({\boldsymbol x})=\sum_{i\in V}f_i({\boldsymbol x})\big\}$ with local objective functions $f_i$ depending only on neighboring variables of the vertex $i\in V$. We introduce a divide-and-conquer algorithm to solve the above optimization problem in a distributed and decentralized manner. The proposed divide-and-conquer algorithm has exponential convergence, its computational cost is almost linear with respect to the size of the network, and it can be fully implemented at fusion centers of the network. Our numerical demonstrations also indicate that the proposed divide-and-conquer algorithm has superior performance than popular decentralized optimization methods do for the least squares problem with/without $\ell^1$ penalty.

math.OC↗
arXiv2021

Polynomial graph filter of multiple shifts and distributed implementation of inverse filtering

Polynomial graph filters and their inverses play important roles in graph signal processing. An advantage of polynomial graph filters is that they can be implemented in a distributed manner, which involves data transmission between adjacent vertices only. The challenge arisen in the inverse filtering is that a direct implementation may suffer from high computational burden, as the inverse graph filter usually has full bandwidth even if the original filter has small bandwidth. In this paper, we consider distributed implementation of the inverse filtering procedure for a polynomial graph filter of multiple shifts, and we propose two iterative approximation algorithms that can be implemented in a distributed network, where each vertex is equipped with systems for limited data storage, computation power and data exchanging facility to its adjacent vertices. We also demonstrate the effectiveness of the proposed iterative approximation algorithms to implement the inverse filtering procedure and their satisfactory performance to denoise time-varying graph signals and a data set of US hourly temperature at 218 locations.

cs.IT↗
arXiv2021

Masked GANs for Unsupervised Depth and Pose Prediction with Scale Consistency

Previous work has shown that adversarial learning can be used for unsupervised monocular depth and visual odometry (VO) estimation, in which the adversarial loss and the geometric image reconstruction loss are utilized as the mainly supervisory signals to train the whole unsupervised framework. However, the performance of the adversarial framework and image reconstruction is usually limited by occlusions and the visual field changes between frames. This paper proposes a masked generative adversarial network (GAN) for unsupervised monocular depth and ego-motion estimation.The MaskNet and Boolean mask scheme are designed in this framework to eliminate the effects of occlusions and impacts of visual field changes on the reconstruction loss and adversarial loss, respectively. Furthermore, we also consider the scale consistency of our pose network by utilizing a new scale-consistency loss, and therefore, our pose network is capable of providing the full camera trajectory over a long monocular sequence. Extensive experiments on the KITTI dataset show that each component proposed in this paper contributes to the performance, and both our depth and trajectory predictions achieve competitive performance on the KITTI and Make3D datasets.

cs.CV↗
arXiv2021

Deep Direct Visual Odometry

Traditional monocular direct visual odometry (DVO) is one of the most famous methods to estimate the ego-motion of robots and map environments from images simultaneously. However, DVO heavily relies on high-quality images and accurate initial pose estimation during tracking. With the outstanding performance of deep learning, previous works have shown that deep neural networks can effectively learn 6-DoF (Degree of Freedom) poses between frames from monocular image sequences in the unsupervised manner. However, these unsupervised deep learning-based frameworks cannot accurately generate the full trajectory of a long monocular video because of the scale-inconsistency between each pose. To address this problem, we use several geometric constraints to improve the scale-consistency of the pose network, including improving the previous loss function and proposing a novel scale-to-trajectory constraint for unsupervised training. We call the pose network trained by the proposed novel constraint as TrajNet. In addition, a new DVO architecture, called deep direct sparse odometry (DDSO), is proposed to overcome the drawbacks of the previous direct sparse odometry (DSO) framework by embedding TrajNet. Extensive experiments on the KITTI dataset show that the proposed constraints can effectively improve the scale-consistency of TrajNet when compared with previous unsupervised monocular methods, and integration with TrajNet makes the initialization and tracking of DSO more robust and accurate.

cs.CV↗
arXiv2020

Inertial Proximal ADMM for Separable Multi-Block Convex Optimizations and Compressive Affine Phase Retrieval

Separable multi-block convex optimization problem appears in many mathematical and engineering fields. In the first part of this paper, we propose an inertial proximal ADMM to solve a linearly constrained separable multi-block convex optimization problem, and we show that the proposed inertial proximal ADMM has global convergence under mild assumptions on the regularization matrices. Affine phase retrieval arises in holography, data separation and phaseless sampling, and it is also considered as a nonhomogeneous version of phase retrieval that has received considerable attention in recent years. Inspired by convex relaxation of vector sparsity and matrix rank in compressive sensing and by phase lifting in phase retrieval, in the second part of this paper, we introduce a compressive affine phase retrieval via lifting approach to connect affine phase retrieval with multi-block convex optimization, and then based on the proposed inertial proximal ADMM for multi-block convex optimization, we propose an algorithm to recover sparse real signals from their (noisy) affine quadratic measurements. Our numerical simulations show that the proposed algorithm has satisfactory performance for affine phase retrieval of sparse real signals.

math.NA↗
arXiv2020

Multi-task GANs for Semantic Segmentation and Depth Completion with Cycle Consistency

Semantic segmentation and depth completion are two challenging tasks in scene understanding, and they are widely used in robotics and autonomous driving. Although several works are proposed to jointly train these two tasks using some small modifications, like changing the last layer, the result of one task is not utilized to improve the performance of the other one despite that there are some similarities between these two tasks. In this paper, we propose multi-task generative adversarial networks (Multi-task GANs), which are not only competent in semantic segmentation and depth completion, but also improve the accuracy of depth completion through generated semantic images. In addition, we improve the details of generated semantic images based on CycleGAN by introducing multi-scale spatial pooling blocks and the structural similarity reconstruction loss. Furthermore, considering the inner consistency between semantic and geometric structures, we develop a semantic-guided smoothness loss to improve depth completion results. Extensive experiments on Cityscapes dataset and KITTI depth completion benchmark show that the Multi-task GANs are capable of achieving competitive performance for both semantic segmentation and depth completion tasks.

cs.CV↗
arXiv2020

Distributed algorithms to determine eigenvectors of matrices on spatially distributed networks

Eigenvectors of matrices on a network have been used for understanding spectral clustering and influence of a vertex. For matrices with small geodesic-width, we propose a distributed iterative algorithm in this letter to find eigenvectors associated with their given eigenvalues. We also consider the implementation of the proposed algorithm at the vertex/agent level in a spatially distributed network.

cs.DC↗
arXiv2020

Preconditioned Gradient Descent Algorithm for Inverse Filtering on Spatially Distributed Networks

Graph filters and their inverses have been widely used in denoising, smoothing, sampling, interpolating and learning. Implementation of an inverse filtering procedure on spatially distributed networks (SDNs) is a remarkable challenge, as each agent on an SDN is equipped with a data processing subsystem with limited capacity and a communication subsystem with confined range due to engineering limitations. In this letter, we introduce a preconditioned gradient descent algorithm to implement the inverse filtering procedure associated with a graph filter having small geodesic-width. The proposed algorithm converges exponentially, and it can be implemented at vertex level and applied to time-varying inverse filtering on SDNs.

cs.IT↗
arXiv2020

Monocular Depth Estimation Based On Deep Learning: An Overview

Depth information is important for autonomous systems to perceive environments and estimate their own state. Traditional depth estimation methods, like structure from motion and stereo vision matching, are built on feature correspondences of multiple viewpoints. Meanwhile, the predicted depth maps are sparse. Inferring depth information from a single image (monocular depth estimation) is an ill-posed problem. With the rapid development of deep neural networks, monocular depth estimation based on deep learning has been widely studied recently and achieved promising performance in accuracy. Meanwhile, dense depth maps are estimated from single images by deep neural networks in an end-to-end manner. In order to improve the accuracy of depth estimation, different kinds of network frameworks, loss functions and training strategies are proposed subsequently. Therefore, we survey the current monocular depth estimation methods based on deep learning in this review. Initially, we conclude several widely used datasets and evaluation indicators in deep learning-based depth estimation. Furthermore, we review some representative existing methods according to different training manners: supervised, unsupervised and semi-supervised. Finally, we discuss the challenges and provide some ideas for future researches in monocular depth estimation.

cs.CV↗
arXiv2020

Random sampling and reconstruction of concentrated signals in a reproducing kernel space

In this paper, we consider (random) sampling of signals concentrated on a bounded Corkscrew domain $Ω$ of a metric measure space, and reconstructing concentrated signals approximately from their (un)corrupted sampling data taken on a sampling set contained in $Ω$. We establish a weighted stability of bi-Lipschitz type for a (random) sampling scheme on the set of concentrated signals in a reproducing kernel space. The weighted stability of bi-Lipschitz type provides a weak robustness to the sampling scheme, however due to the nonconvexity of the set of concentrated signals, it does not imply the unique signal reconstruction. From (un)corrupted samples taken on a finite sampling set contained in $Ω$, we propose an algorithm to find approximations to signals concentrated on a bounded Corkscrew domain $Ω$. Random sampling is a sampling scheme where sampling positions are randomly taken according to a probability distribution. Next we show that, with high probability, signals concentrated on a bounded Corkscrew domain $Ω$ can be reconstructed approximately from their uncorrupted (or randomly corrupted) samples taken at i.i.d. random positions drawn on $Ω$, provided that the sampling size is at least of the order $μ(Ω) \ln (μ(Ω))$, where $μ(Ω)$ is the measure of the concentrated domain $Ω$. Finally, we demonstrate the performance of proposed approximations to the original concentrated signal when the sampling procedure is taken either with small density or randomly with large size.

cs.IT↗
arXiv2020

When Autonomous Systems Meet Accuracy and Transferability through AI: A Survey

With widespread applications of artificial intelligence (AI), the capabilities of the perception, understanding, decision-making and control for autonomous systems have improved significantly in the past years. When autonomous systems consider the performance of accuracy and transferability, several AI methods, like adversarial learning, reinforcement learning (RL) and meta-learning, show their powerful performance. Here, we review the learning-based approaches in autonomous systems from the perspectives of accuracy and transferability. Accuracy means that a well-trained model shows good results during the testing phase, in which the testing set shares a same task or a data distribution with the training set. Transferability means that when a well-trained model is transferred to other testing domains, the accuracy is still good. Firstly, we introduce some basic concepts of transfer learning and then present some preliminaries of adversarial learning, RL and meta-learning. Secondly, we focus on reviewing the accuracy or transferability or both of them to show the advantages of adversarial learning, like generative adversarial networks (GANs), in typical computer vision tasks in autonomous systems, including image style transfer, image superresolution, image deblurring/dehazing/rain removal, semantic segmentation, depth estimation, pedestrian detection and person re-identification (re-ID). Then, we further review the performance of RL and meta-learning from the aspects of accuracy or transferability or both of them in autonomous systems, involving pedestrian tracking, robot navigation and robotic manipulation. Finally, we discuss several challenges and future topics for using adversarial learning, RL and meta-learning in autonomous systems.

cs.LG↗
arXiv2019

Polynomial control on weighted stability bounds and inversion norms of localized matrices on simple graphs

The (un)weighted stability for some matrices is one of essential hypotheses in time-frequency analysis and applied harmonic analysis. In the first part of this paper, we show that for a localized matrix in a Beurling algebra, its weighted stabilities for different exponents and Muckenhoupt weights are equivalent to each other, and reciprocal of its optimal lower stability bound for one exponent and weight is controlled by a polynomial of reciprocal of its optimal lower stability bound for another exponent and weights. Inverse-closed Banach subalgebras of matrices with certain off-diagonal decay can be informally interpreted as localization preservation under inversion, which is of great importance in many mathematical and engineering fields. Let ${\mathcal B}(\ell^p_w)$ be the Banach algebra of bounded operators on the weighted sequence space $\ell^p_w$ on a simple graph. In the second part of this paper, we prove that Beurling algebras of localized matrices on a simple graph are inverse-closed in ${\mathcal B}(\ell^p_w)$ for all $1\le p<\infty$ and Muckenhoupt $A_p$-weights $w$, and the Beurling norm of the inversion of a matrix $A$ is bounded by a bivariate polynomial of the Beurling norm of the matrix $A$ and the operator norm of its inverse $A^{-1}$ in ${\mathcal B}(\ell^p_w)$.

math.FA↗
arXiv2019

Phase retrieval of complex and vector-valued functions

The phase retrieval problem in the classical setting is to reconstruct real/complex functions from the magnitudes of their Fourier/frame measurements. In this paper, we consider a new phase retrieval paradigm in the complex/quaternion/vector-valued setting, and we provide several characterizations to determine complex/quaternion/vector-valued functions $f$ in a linear space $S$ of (in)finite dimensions, up to a trivial ambiguity, from the magnitudes $\|ϕ(f)\|$ of their linear measurements $ϕ(f), ϕ\in Φ$. Our characterization in the scalar setting implies the well-known equivalence between the complement property for linear measurements $Φ$ and the phase retrieval of linear space $S$. In this paper, we also discuss the affine phase retrieval of vector-valued functions in a linear space and the reconstruction of vector fields on a graph, up to an orthogonal matrix, from their absolute magnitudes at vertices and relative magnitudes between neighboring vertices.

math.FA↗
arXiv2019

Space-Time Sampling for Network Observability

Designing sparse sampling strategies is one of the important components in having resilient estimation and control in networked systems as they make network design problems more cost-effective due to their reduced sampling requirements and less fragile to where and when samples are collected. It is shown that under what conditions taking coarse samples from a network will contain the same amount of information as a more finer set of samples. Our goal is to estimate initial condition of linear time-invariant networks using a set of noisy measurements. The observability condition is reformulated as the frame condition, where one can easily trace location and time stamps of each sample. We compare estimation quality of various sampling strategies using estimation measures, which depend on spectrum of the corresponding frame operators. Using properties of the minimal polynomial of the state matrix, deterministic and randomized methods are suggested to construct observability frames. Intrinsic tradeoffs assert that collecting samples from fewer subsystems dictates taking more samples (in average) per subsystem. Three scalable algorithms are developed to generate sparse space-time sampling strategies with explicit error bounds.

eess.SY↗
arXiv2019

Koopman Performance Analysis of Nonlinear Consensus Networks

Spectral decomposition of dynamical systems is a popular methodology to investigate the fundamental qualitative and quantitative properties of these systems and their solutions. In this chapter, we consider a class of nonlinear cooperative protocols, which consist of multiple agents that are coupled together via an undirected state-dependent graph. We develop a representation of the system solution by decomposing the nonlinear system utilizing ideas from the Koopman operator theory and its spectral analysis. We use recent results on the extensions of the well-known Hartman theorem for hyperbolic systems to establish a connection between the original nonlinear dynamics and the linearized dynamics in terms of Koopman spectral properties. The expected value of the output energy of the nonlinear protocol, which is related to the notions of coherence and robustness in dynamical networks, is evaluated and characterized in terms of Koopman eigenvalues, eigenfunctions, and modes. Spectral representation of the performance measure enables us to develop algorithmic methods to assess the performance of this class of nonlinear dynamical networks as a function of their graph topology. Finally, we propose a scalable computational method for approximation of the components of the Koopman mode decomposition, which is necessary to evaluate the systemic performance measure of the nonlinear dynamic network.

eess.SY↗