To address the problem that task-irrelevant objects such as cars, pedestrians and sky, will interfere with the extracted feature descriptors in cross-view image geo-localization, this paper proposes a novel method for cross-view image geo-localization, named as AENet. The method includes two main parts: an attention efficient network fusing channel and spatial attention mechanisms and a triplet loss function based on a multiple hard samples weighting strategy. In the first part, the EfficientNetV2 network is used to extract features from the images and preliminarily filter irrelevant features from the channel dimension, then the Triplet Attention layer is applied to further filter irrelevant features from the spatial dimension. In the second part, a multiple hard samples weighting strategy is proposed to enhance the learning of hard samples. Experimental results show that our proposed method significantly outperforms the state-of-the-art method on two existing benchmark datasets.
Citation: Jingqian Xu, Ma Zhu, Baojun Qi, Jiangshan Li, Chunfang Yang. AENet: attention efficient network for cross-view image geo-localization[J]. Electronic Research Archive, 2023, 31(7): 4119-4138. doi: 10.3934/era.2023210
[1] | Xintao Li, Rongrui Lin, Lianbing She . Periodic measures for a neural field lattice model with state dependent superlinear noise. Electronic Research Archive, 2024, 32(6): 4011-4024. doi: 10.3934/era.2024180 |
[2] | Shuang Wang, FanFan Chen, Chunlian Liu . The existence of periodic solutions for nonconservative superlinear second order ODEs: a rotation number and spiral analysis approach. Electronic Research Archive, 2025, 33(1): 50-67. doi: 10.3934/era.2025003 |
[3] | Lianbing She, Nan Liu, Xin Li, Renhai Wang . Three types of weak pullback attractors for lattice pseudo-parabolic equations driven by locally Lipschitz noise. Electronic Research Archive, 2021, 29(5): 3097-3119. doi: 10.3934/era.2021028 |
[4] | Nan Xiang, Aying Wan, Hongyan Lin . Diffusion-driven instability of both the equilibrium solution and the periodic solutions for the diffusive Sporns-Seelig model. Electronic Research Archive, 2022, 30(3): 813-829. doi: 10.3934/era.2022043 |
[5] | Zhili Zhang, Aying Wan, Hongyan Lin . Spatiotemporal patterns and multiple bifurcations of a reaction- diffusion model for hair follicle spacing. Electronic Research Archive, 2023, 31(4): 1922-1947. doi: 10.3934/era.2023099 |
[6] | Yixuan Wang, Xianjiu Huang . Ground states of Nehari-Pohožaev type for a quasilinear Schrödinger system with superlinear reaction. Electronic Research Archive, 2023, 31(4): 2071-2094. doi: 10.3934/era.2023106 |
[7] | Xiangwen Yin . A review of dynamics analysis of neural networks and applications in creation psychology. Electronic Research Archive, 2023, 31(5): 2595-2625. doi: 10.3934/era.2023132 |
[8] | Weiyu Li, Hongyan Wang . Dynamics of a three-molecule autocatalytic Schnakenberg model with cross-diffusion: Turing patterns of spatially homogeneous Hopf bifurcating periodic solutions. Electronic Research Archive, 2023, 31(7): 4139-4154. doi: 10.3934/era.2023211 |
[9] | Peng Yu, Shuping Tan, Jin Guo, Yong Song . Data-driven optimal controller design for sub-satellite deployment of tethered satellite system. Electronic Research Archive, 2024, 32(1): 505-522. doi: 10.3934/era.2024025 |
[10] | Peng Gao, Pengyu Chen . Blowup and MLUH stability of time-space fractional reaction-diffusion equations. Electronic Research Archive, 2022, 30(9): 3351-3361. doi: 10.3934/era.2022170 |
To address the problem that task-irrelevant objects such as cars, pedestrians and sky, will interfere with the extracted feature descriptors in cross-view image geo-localization, this paper proposes a novel method for cross-view image geo-localization, named as AENet. The method includes two main parts: an attention efficient network fusing channel and spatial attention mechanisms and a triplet loss function based on a multiple hard samples weighting strategy. In the first part, the EfficientNetV2 network is used to extract features from the images and preliminarily filter irrelevant features from the channel dimension, then the Triplet Attention layer is applied to further filter irrelevant features from the spatial dimension. In the second part, a multiple hard samples weighting strategy is proposed to enhance the learning of hard samples. Experimental results show that our proposed method significantly outperforms the state-of-the-art method on two existing benchmark datasets.
This paper deals with periodic measures of the following reaction-diffusion lattice systems driven by superlinear noise defined on the integer set Zk :
dui(t)+λ(t)ui(t)dt−ν(t)(u(i1−1,i2,…,ik)(t)+ui1,i2−1,…,ik(t)+…+ui1,i2,…,ik−1(t)−2ku(i1,i2,…,ik)(t)+u(i1+1,i2,…,ik)(t)+u(i1,i2+1,…,ik)(t)+…+u(i1,i2,…,ik+1)(t))dt=fi(t,ui(t))dt+gi(t)dt+∞∑j=1(hi,j(t)+δi,jˆσi,j(t,ui(t)))dWj(t), | (1.1) |
along with initial conditions:
ui(0)=u0,i, | (1.2) |
where i=(i1,i2,…,ik)∈Zk, λ(t),ν(t) are continuous functions, λ(t)>0, (fi)i∈Zk and (ˆσi,j)i∈Zk,j∈N are two sequences of continuously differentiable nonlinearities with arbitrary and superlinear growth rate from R×R→R, respectively, g=(gi)i∈Zk and h=(hi,j)i∈Zk,j∈N are two time-dependent random sequences, and δ=(δi,j)i∈Zk,j∈N is a sequence of real numbers. The sequence of independent two-sided real-valued Wiener processes (Wj)j∈N is defined on a complete filtered probability space (Ω,F,{Ft}t∈R,P). Furthermore, we assume that system (1.1) is a time periodic system; more precisely, there exists T>0 such that the time-dependent functions λ,ν,fi,g,h,σi,j(i∈Zk,j∈N) in (1.1) are all T-periodic in time.
Lattice systems are gradually becoming a large and evolving interdisciplinary research field, due to wide range of applications in physics, biology and engineering such as pattern recognition, propagation of nerve pulses, electric circuits, and so on, see [1,2,3,4,5,6] and the references therein for more details. The well-posedness and the dynamics of these equations have been studied by many authors, [7,8,9,10] for deterministic systems and [11,12,13,14,15,16,17,18,19] for stochastic systems where the existence of random attractors and probability measures have been examined. Especially, the authors research the limiting behavior of periodic measures of lattice systems in [15].
Nonlinear noise was proposed and studied for the first time in [19], the authors researches the long-term behavior of lattice systems driven by nonlinear noise in terms of random attractors and invariant measures. Before that, the research on noise was limited to additive noise and linear multiplicative noise, which can be transformed into a deterministic system. However, if the diffusion coefficients are nonlinear, then one cannot convert the stochastic system into a pathwise deterministic one, and thereby this problem cannot be studied under the frameworks of deterministic systems aforementioned. As an extension of [19], a class of reaction-diffusion lattice systems driven by superlinear noise, where the noise has a superlinear growth order q∈[2,p), is studied by taking advantage of the dissipativeness of the nonlinear drift function fi in (1.1) to control the superlinear noise in [20].
In the paper, we will study the existence of periodic measures of reaction-diffusion lattice systems drive by superlinear noise. One of the main tasks in our analysis is to solve the superlinear noise terms. We remark that if the noise grows linearly, then the estimates we need can be obtained by applying the standard methods available in the literature. We adopt the ideas that take advantage of the nonlinear drift terms' the polynomical growth rate p (p≥2) to control the noise polynomical rate q∈[2,p). Furthermore, notice that l2 is an infinite-dimensional phase space and problem (1.1)–(1.2) is defined on the unbounded set Zk. The unboundedness of Zk as well as the infinite-dimensionalness of l2 introduce a major difficulty, because of the non-compactness of usual Sobolev embeddings on unbounded domains. We will employ the dissipativeness of the drift function in (1.1) as well as a cutoff technique to prove that the tails of solutions are uniformly small in L2(Ω,l2). Based upon this fact we obtain the tightness of distribution laws of solutions, and then the existence of periodic measures.
In the next section, we discuss the well-poseness of solutions of (1.1) and (1.2). Section 3 is devoted to the uniform estimates of solutions including the uniform estimates on the tails of solutions. In Section 4, we show the existence of periodic measures of (1.1) and (1.2).
In this section, we prove the existence and uniqueness of solutions to system (1.1) and (1.2). We first discuss the assumptions on the nonlinear drift and diffusion terms in (1.1).
We begin with the following Banach space:
lr={u=(ui)i∈Zk:∑i∈Zk|ui|r<+∞} with norm ‖u‖r=(∑i∈Zk|ui|r)1r,∀r≥1. |
The norm and inner product of l2 are denoted by (⋅,⋅) and ‖⋅‖, respectively. For the nonlinear drift function fi∈C1(R×R,R) in the equation we assume that for all s∈R and i∈Zk,
fi(t,s)s≤−γ1|s|p+ϕ1,i, ϕ1={ϕ1,i}i∈Zk∈l1, | (2.1) |
|fi(t,s)|≤ϕ2,i|s|p−1+ϕ3,i, ϕ2={ϕ2,i}i∈Zk∈l∞, ϕ3={ϕ3,i}i∈Zk∈l2, | (2.2) |
|f′i(t,s)|≤ϕ4,i|s|p−2+ϕ5,i, ϕ4={ϕ4,i}i∈Zk∈l∞, ϕ5={ϕ5,i}i∈Zk∈l∞, | (2.3) |
where p>2 and γ1>0 are constants. For the sequence of continuously differentiable diffusion functions ˆσ=(ˆσi,j)i∈Zk,j∈N, we assume, for all s∈R and j∈N,
|ˆσi,j(t,s)|≤φ1,i|s|q2+φ2,i, φ1={φ1,i}i∈Zk∈l2pp−q, φ2={φ2,i}i∈Zk∈l2, | (2.4) |
|ˆσ′i,j(t,s)|≤φ3,i|s|q2−1+φ4,i, φ3={φ3,i}i∈Zk∈lq, φ4={φ4,i}i∈Zk∈l∞, | (2.5) |
where q∈[2,p) is a constant. For processes g(t)=(gi(t))i∈Zk and h(t)=(hi,j)i∈Zk,j∈N are both continuous in t∈R, which implies that for all t∈R,
‖g(t)‖2=∑i∈Zk|gi(t)|2<∞ and ‖h(t)‖2=∑i∈Zk∑j∈N|hi,j(t)|2<∞. | (2.6) |
In addition, we assume δ=(δi,j)i∈Zk,j∈N satisfies
cδ:=∑j∈N∑i∈Zk|δi,j|2<∞. | (2.7) |
We will investigate the periodic measures of system (1.1)–(1.2) for which we assume that all given time-dependent functions are T-periodic in t∈R for some T>0; that is, for all t∈R,i∈Zk and k∈N.
λ(t+T)=λ(t),ν(t+T)=ν(t),h(t+T)=h(t),g(t+T)=g(t),f(t+T,⋅)=f(t,⋅),σ(t+T,⋅)=σ(t,⋅). |
If m:R→R is a continuous T-periodic function, we denote
¯m=max0≤t≤Tm(t),m_=min0≤t≤Tm(t). |
We want to reformulate problem (1.1)–(1.2) as an abstract one in l2. Given 1≤j≤k,u=(ui)i∈Zk∈l2 and i=(i1,i2,…,ik)∈Zk. Let us define the operators from l2 to l2 by
(Bju)i=u(i1,…,ij+1,…,ik)−u(i1,…,ij,…,ik),(B∗ju)i=u(i1,…,ij−1,…,ik)−u(i1,…,ij,…,ik),(Aju)i=−u(i1,…,ij+1,…,ik)+2u(i1,…,ij,…,ik)−u(i1,…,ij−1,…,ik), |
and
(Aku)i=−u(i1−1,i2,…,ik)−u(i1,i2−1,…,ik)−…−u(i1,i2,…,ik−1)+2ku(i1,i2,…,ik)−u(i1+1,i2,…,ik)−u(i1,i2+1,…,ik)−…−u(i1,i2,…,ik+1). |
For all 1≤j≤k,u=(ui)i∈Zk∈l2 and v=(vi)i∈Zk∈l2 we see
‖Bju‖≤2‖u‖,(B∗ju,v)=(u,Bjv),Aj=BjB∗j and Ak=k∑j=1Aj. | (2.8) |
Again, define the operators f,σj:R×l2→l2 by
f(t,u)=(fi(t,ui))i∈Zk and σj(t,u)=(δi,jˆσi,j(t,ui))i∈Zk,∀t∈R,∀u=(ui)i∈Zk∈l2. |
It follows from (2.3) that there exists θ∈(0,1) such that for p>2 and u,v∈l2,
∑i∈Zk|fi(t,ui)−fi(t,vi)|2=∑i∈Zk|f′i(θui+(1−θ)vi)|2|ui−vi|2≤∑i∈Zk(|ϕ4,i||θui+(1−θ)vi|p−2+|ϕ5,i|)2|ui−vi|2≤∑i∈Zk(22p−4|ϕ4,i|2(|ui|2p−4+|vi|2p−4)+2|ϕ5,i|2)|ui−vi|2≤(22p−4‖ϕ4‖2l∞(‖u‖2p−4+‖v‖2p−4)+2‖ϕ5‖2l∞)‖u−v‖2. | (2.9) |
This together with f(t,0)∈l2 by (2.2) yields f(t,u)∈l2 for all u∈l2, and thereby f:R×l2→l2 is well-defined. In addition, we deduce from (2.9) that f:R×l2→l2 is a locally Lipschitz continuous function, that is, for every n∈N, we can find a constant c1(n)>0 satisfying, for all u,v∈l2 with ‖u‖≤n and ‖v‖≤n,
‖f(u)−f(v)‖≤c1(n)‖u−v‖. | (2.10) |
For q∈[2,p) and u∈l2, one can deduce from(2.4), (2.7) and Young's inequality that for all ϖ>0,
ϖ∑j∈N‖σj(t,u)‖2=ϖ∑j∈N∑i∈Zk|δi,jˆσi,j(t,ui)|2≤2ϖ∑j∈N∑i∈Zk|δi,j|2(|φ1,i|2|ui|q+|φ2,i|2)≤2ϖcδ∑i∈Zk(|φ1,i|2|ui|q+|φ2,i|2)≤γ12∑i∈Zk|ui|p+p−qp(pγ12q)−qp−q(2ϖcδ)pp−q∑i∈Zk|φ1,i|2pp−q+2ϖcδ∑i∈Zk|φ2,i|2≤γ12‖u‖pp+p−qp(pγ12q)−qp−q(2ϖcδ)pp−q‖φ1‖2pp−q2pp−q+2ϖcδ‖φ2‖2, | (2.11) |
where γ1 is the same number as in (2.1). From (2.11) and l2⊆lp for p>2, we find that σj(t,u)∈l2 for all u∈l2. Then σj:R×l2→l2 is also well-defined. In addition, it yields from (2.5) and (2.7) that there exists η∈(0,1) such that for q∈[2,p) and u,v∈l2,
∑j∈N∑i∈Zk|δi,jˆσi,j(t,ui)−δi,jˆσi,j(t,vi)|2=∑i∈Zk∑j∈N|δi,j|2|ˆσi,j(t,ui)−ˆσi,j(t,vi)|2=∑i∈Zk∑j∈N|δi,j|2|ˆσ′i,j(ηui+(1−η)vi)|2|ui−vi|2≤cδ∑i∈Zk(|φ3,i||ηui+(1−η)vi|q2−1+|φ4,i|)2|ui−vi|2≤cδ∑i∈Zk(2q−2|φ3,i|2(|ui|q−2+|vi|q−2)+2|φ4,i|2)|ui−vi|2≤cδ∑i∈Zk(2q−2(4q|φ3,i|q+q−2q|ui|q+q−2q|vi|q)+2|φ4,i|2)|ui−vi|2≤cδ(2q−1(‖φ3‖qq+‖u‖q+‖v‖q)+2‖φ4‖2l∞)‖u−v‖2. | (2.12) |
This implies that σj:R×l2→l2 is also locally Lipschitz continuous, more precisely, for every n∈N, one can find a constant c2(n)>0 satisfying, for all u,v∈l2 with ‖u‖≤n and ‖v‖≤n,
∑j∈N‖σj(u)‖2≤c22(n). | (2.13) |
and
∑j∈N‖σj(u)−σj(v)‖2≤c22(n)‖u−v‖2. | (2.14) |
By above notations one is able to rewrite (1.1)–(1.2) as the following system in l2 for t>0 :
du(t)+ν(t)Aku(t)dt+λ(t)u(t)dt=f(t,u(t))dt+g(t)dt+∞∑j=1(hj(t)+σj(t,u(t)))dWj(t), | (2.15) |
with initial condition:
u(0)=u0∈l2, | (2.16) |
in the present article, the solutions of system (2.15)–(2.16) are interpreted in the following sense.
Definition 2.1. Suppose u0∈L2(Ω,l2) is F0-measurable, a continuous l2-valued Ft-adapted stochastic process u is called a solution of equations (2.15) and (2.16) if u∈L2(Ω,C([0,T],l2))∩Lp(Ω,Lp(0,T;lp)) for all T>0, and the following equation holds for all t≥0 and almost all ω∈Ω:
u(t)=u0+∫t0(−ν(s)Aku(s)−λ(s)u(s)+f(s,u(s))+g(s))ds+∞∑j=1∫t0(hj(s)+σj(s,u(s)))dWj(s) in l2. | (2.17) |
Similar to Ref.[20], we can get (2.15) and (2.16) exist global solutions in the sense of Definition 2.1.
In this section, we derive the uniform estimates of solutions of (2.15)–(2.16). These estimates will be used to establish the tightness of a set of probability distributions of u in l2.
We assume that
α(t)=λ(t)−16k|ν(t)|>0. | (3.1) |
Lemma 3.1. Let (2.1)–(2.7) and (3.1) hold. Then the solutions u(t,0,u0) of system (2.15) and (2.16) with initial data u0 at time 0 satisfy, for all t≥0,
E(‖u(t,0,u0)‖2)+∫t0eα_(r−t)E(‖u(r,0,u0)‖pp)dr≤L1(E(‖u0‖2)+∞∑j=1¯‖hj‖2+¯‖g‖2+‖φ1‖2pp−q2pp−q+‖φ2‖2+‖ϕ1‖1), | (3.2) |
where L1>0 is a positive constant which depends on α_,p,q,γ,cδ,t, but indepentent of u0.
Proof. Applying Ito's formula to (2.15) we get
d(‖u(t)‖2)+2ν(t)k∑j=1‖Bju(t)‖2dt+2λ(t)‖u(t)‖2dt=2(f(t,u(t)),u(t))dt+2(g(t),u(t))dt+∞∑j=1‖hj(t)+σ(t,u(t))‖2dt+2∞∑j=1u(t)(hj(t)+σj(t,u(t)))dWj(t). |
This implies
ddtE(‖u(t)‖2)+2ν(t)k∑j=1E(‖Bju(t)‖2)+2λ(t)E(‖u(t)‖2)≤2E(f(t,u(t)),u(t))+2E(g(t),u(t))+2∞∑j=1E(‖hj(t)‖2)+2∞∑j=1E(‖σ(t,u(t))‖2). | (3.3) |
For the second term on the left-hand side of (3.3), we have
2|ν(t)|k∑j=1E(‖Bju(t)‖2)≤8k|ν(t)|E(‖u(t)‖2). | (3.4) |
For the first term on the right-hand side of (3.3), we get from (2.1) that
2E(f(t,u(t)),u(t))≤−2γ1E(‖u(t)‖pp)+2‖ϕ1‖1. | (3.5) |
For the second term on the right-hand side of (3.3), we have
2E(g(t),u(t))≤λ(t)E(‖u(t)‖2)+1λ(t)E(‖g(t)‖2). | (3.6) |
For the last term on the right-hand side of (3.3), we infer from (2.11) with ω=2 that
2∞∑j=1E(‖σj(t,u(t))‖2)≤γ12E(‖u(t)‖pp)+p−qp(pγ12q)−qp−q(4cδ)pp−q‖φ1‖2pp−q2pp−q+4cδ‖φ2‖2. | (3.7) |
By (3.3)–(3.7) we get
ddtE(‖u(t)‖2)+α_E(‖u(t)‖2)+32γ1E(‖u(t)‖pp)≤E(∞∑j=12‖hj(t)‖2+1λ(t)‖g(t)‖2)+C1, | (3.8) |
implies that
ddtE(‖u(t)‖2)+α_E(‖u(t)‖2)+32γ1E(‖u(t)‖pp)≤2∞∑j=1‖¯hj‖2+1λ_‖¯g‖2+C1, | (3.9) |
where C1=p−qp(pγ12q)−qp−q(4cδ)pp−q‖φ1‖2pp−q2pp−q+4cδ‖φ2‖2+2‖ϕ1‖1. Multiplying (3.9) by eα_t and integrating over (0,t) to obtain
E(‖u(t,0,u0)‖2)+32γ1∫t0eα_(r−t)E(‖u(r,0,u0)‖pp)dr≤e−α_tE(‖u0‖2)+C2∫t0eα_(r−t)dr, | (3.10) |
where C2=2∑∞j=1‖¯hj‖2+1λ_‖¯g‖2+C1. This completes the proof.
Lemma 3.2. Let (2.1)–(2.7), and (3.1) be satisfied. Then for compact subset K of l2, one can find a number N0=N0(K)∈N such that the solutions u(t,0,u0) of (2.15) and (2.16) satisfy, for all n≥N0 and t≥0,
E(∑‖i‖≥n|ui(t,0,u0)|2)+∫t0eα_(r−t)E(∑‖i‖≥n|ui(r,0,u0)|p)dr≤ε, | (3.11) |
where u0∈K and ‖i‖:=maxi≤j≤k|ij|.
Proof. Define a smooth function ξ:R→[0,1] such that
ξ(s)=0 for |s|≤1 and ξ(s)=1 for |s|≥2. | (3.12) |
Denote by
ξn=(ξ(‖i‖n))i∈Zk and ξnu=(ξ(‖i‖n)ui)i∈Zk,∀u=(ui)i∈Zk,n∈N. | (3.13) |
Similar notations will also be used for other terms. It follows from (2.15) that
d(ξnu(t))+ν(t)ξnAku(t)dt+λ(t)ξnu(t)dt=ξnf(t,u(t))dt+ξng(t)dt+∞∑j=1(ξnhj(t)+ξnσj(t,u(t)))dWj(t). | (3.14) |
By Ito's formula and (3.14) we have
d‖ξnu(t)‖2+2ν(t)(Ak(u(t)),ξ2nu(t))dt+2λ(t)‖ξnu(t)‖2dt=2(f(t,u(t)),ξ2nu(t))dt+2(g(t),ξ2nu(t))dt+∞∑j=1‖ξnhj(t)+ξnσj(t,u(t))‖2dt+2∞∑j=1(hj(t)+σj(t,u(t)),ξ2nu(t))dWj. | (3.15) |
This yields
ddtE(‖ξnu(t)‖2)+2ν(t)E(Ak(u(t)),ξ2nu(t))+2λ(t)E(‖ξnu(t)‖2)=2E(f(t,u(t)),ξ2nu(t))+2E(g(t),ξ2nu(t))+2∞∑j=1E(‖ξnhj(t)‖2)+2∞∑j=1E(‖ξnσj(t,u(t))‖2)dt. | (3.16) |
For the second term on the left-hand side of (3.16), we have
2ν(t)E(Ak(u(t)),ξ2nu(t))=2ν(t)k∑j=1E(Bju(t),Bj(ξ2nu(t)))=2ν(t)E(k∑j=1∑i∈Zk(ui1,…,ij+1,…,ik−ui)×(ξ2(‖(i1,…,ij+1,…,ik)‖n)u(i1,…,ij+1,…,ik)−ξ2(‖i‖n)ui))=2ν(t)E(k∑j=1∑i∈Zkξ2(‖i‖n)(ui1,…,ij+1,…,ik−ui)2)+2ν(t)E(k∑j=1∑i∈Zk(ξ2(‖(i1,…,ij+1,…,ik)‖n)−ξ2(‖i‖n))×(u(i1,…,ij+1,…,ik)−ui)u(i1,…,ij+1,…,ik)). | (3.17) |
We first deal with the first term on the right-hand side of (3.17). Notice that
2|ν(t)|E(k∑j=1∑i∈Zkξ2(‖i‖n)(ui1,…,ij+1,…,ik−ui)2)=2|ν(t)|E(k∑j=1∑i∈Zk|ξ(‖i‖n)u(i1,…,ij+1,…,ik)−ξ(‖i‖n)ui|2)≤4|ν(t)|E(k∑j=1∑i∈Zk|(ξ(‖i‖n)−ξ(‖(i1,…,ij+1,…,ik)‖n))u(i1,…,ij+1,…,ik)|2)+4|ν(t)|E(k∑j=1∑i∈Zk|ξ(‖(i1,…,ij+1,…,ik)‖n)u(i1,…,ij+1,…,ik)−ξ(‖i‖n)ui|2). | (3.18) |
By the definition of function ξ, there exists a constant C3>0 such that |ξ′(s)|≤C3 for all s∈R. Then the first term on the right-hand side of (3.18) is bounded by
4|ν(t)|E(k∑j=1∑i∈Zk|(ξ(‖i‖n)−ξ(‖(i1,…,ij+1,…,ik)‖n))u(i1,…,ij+1,…,ik)|2)=4|ν(t)|E(k∑j=1∑i∈Zk|ξ(‖i‖n)−ξ(‖(i1,…,ij+1,…,ik)‖n)|2|u(i1,…,ij+1,…,ik)|2)≤4C23n2|ν(t)|E(k∑j=1∑i∈Zk|u(i1,…,ij+1,…,ik)|2)≤4C23kn2|ν(t)|E(‖u‖2). | (3.19) |
By the definition of |Bju|i, the last term on the right-hand side of (3.18) is bounded by
4|ν(t)|E(k∑j=1∑i∈Zk|ξ(‖(i1,…,ij+1,…,ik)‖n)u(i1,…,ij+1,…,ik)−ξ(‖i‖n)ui|2)≤4|ν(t)|E(k∑j=1‖Bj(ξnu(t))‖2)≤16k|ν(t)|E(‖ξnu(t)‖2). | (3.20) |
Then we find from (3.18) to (3.20) that the first term on the right-hand side of (3.17) is bounded by
2|ν(t)|E(k∑j=1∑i∈Zkξ2(‖i‖n)(u(i1,…,ij+1,…,ik)−ui)2)≤16k|ν(t)|E(‖ξnu(t)‖2)+4C23kn2|ν(t)|E(‖u‖2). | (3.21) |
In addition, we find that the last term on the right-hand side of (3.17) can be bounded by
2|ν(t)E(k∑j=1∑i∈Zk(ξ2(‖(i1,…,ij+1,…,ik)‖n)−ξ2(‖i‖n))×(u(i1,…,ij+1,…,ik)−ui)u(i1,…,ij+1,…,ik))|≤2|ν(t)|E(k∑j=1∑i∈Zk|ξ2(‖(i1,…,ij+1,…,ik)‖n)−ξ2(‖i‖n)|×|u(i1,…,ij+1,…,ik)−ui||u(i1,…,ij+1,…,ik)|)≤4|ν(t)|E(k∑j=1∑i∈Zk|ξ(‖(i1,…,ij+1,…,ik)‖n)−ξ(‖i‖n)|×|u(i1,…,ij+1,…,ik)−ui||u(i1,…,ij+1,…,ik)|)≤4C3n|ν(t)|E(k∑j=1∑i∈Zk|u(i1,…,ij+1,…,ik)−ui||u(i1,…,ij+1,…,ik)|)≤8kC3n|ν(t)|E(‖u‖2). | (3.22) |
By (3.21), (3.22) and (3.17), we infer that the second term on the left-hand side of (3.16) satisfied
2|ν(t)E(Ak(u(t)),ξ2nu(t))|≤C4|ν(t)|(1n+1n2)E(‖u‖2)+16k|ν(t)|E(‖ξnu(t)‖2), | (3.23) |
where C4=4kC3(2+C3). For the first term on the right-hand side of (3.16), we find from (2.1) that
2E(f(t,u(t)),ξ2nu(t))≤−2γ1E(∑i∈Zkξ2(‖i‖n)|ui(t)|p)+2E(∑i∈Zkξ2(‖i‖n)|ϕ1,i|)≤−2γ1E(∑i∈Zkξ2(‖i‖n)|ui(t)|p)+2∑‖i‖≥n|ϕ1,i|. | (3.24) |
For the second term on the right-hand side of (3.16), we infer from Young's inequality that
2E(g,ξ2nu(t))≤λ_E(‖ξnu(t)‖2)+1λ_E(∑i∈Zkξ2(‖i‖n)|gi|2)≤λ_E(‖ξnu(t)‖2)+1λ_∑‖i‖≥n|gi|2. | (3.25) |
For the last term on the right-hand side (3.16), we infer from (2.4) and Young's inequality that
2∞∑j=1E(‖ξnσj(t,u(t))‖2)=2∞∑j=1E(∑i∈Zk|ξ(‖i‖n)δi,jˆσi,j(t,ui(t))|2)≤4∞∑j=1E(∑i∈Zkξ2(‖i‖n)|δi,j|2(|φ1,i|2|ui(t)|q+|φ2,i|2))≤4cδE(∑i∈Zkξ2(‖i‖n)(|φ1,i|2|ui(t)|q+|φ2,i|2))≤γ1E(∑i∈Zkξ2(‖i‖n)|ui(t)|p)+p−qp(pγ1q)−qp−q(4cδ)pp−q∑i∈Zkξ2(‖i‖n)|φ1,i|2pp−q+4cδ∑i∈Zkξ2(‖i‖n)|φ2,i|2≤γ1E(∑i∈Zkξ2(‖i‖n)|ui(t)|p)+p−qp(pγ1q)−qp−q(4cδ)pp−q∑‖i‖≥n|φ1,i|2pp−q+4cδ∑‖i‖≥n|φ2,i|2. | (3.26) |
Substituting (3.23)–(3.26) into (3.16) we get
ddtE(‖ξnu(t)‖2)+α_E(‖ξnu(t)‖2)+γ1E(∑i∈Zkξ2(‖i‖n)|ui(t)|p)≤C4|ν|(1n+1n2)E(‖u‖2)+C5(∑‖i‖≥n(¯|gi|2+|φ1,i|2pp−q+|φ2,i|2+|ϕ1,i|)+∑‖i‖≥n∞∑j=1¯|hi,j|2), | (3.27) |
where C5=2+1λ_+p−qp(pγ1q)−qp−q(4cδ)pp−q+4cδ. One can multiply (3.27) by eα_t and integrate over (0,t) in order to obtain
E(‖ξnu(t,0,u0)‖2)+γ1∫t0eα_(r−t)E(∑i∈Zkξ2(‖i‖n)|ui(r,0,u0)|p)dr≤e−α_tE(‖ξnu0‖2)+C4|ν|(1n+1n2)∫t0eα_(r−t)E(‖u(r,0,u0)‖2)dr+C5α_(∑‖i‖≥n(¯|gi|2+|φ1,i|2pp−q+|φ2,i|2+|ϕ1,i|)+∑‖i‖≥n∞∑j=1¯|hi,j|2). | (3.28) |
Since K is a compact subset of l2 we infer from (3.1) that
limn→∞supu0∈Ksupt≥0e−α_tE(‖ξnu0‖2)≤limn→∞supu0∈KE(∑‖i‖≥n|u0,i|2)=0. | (3.29) |
By Lemma 3.1, we find that for all u0∈K and t≥0, as n→∞,
(1n+1n2)∫t0eα_(r−t)E(‖u(r,0,u0)‖2)dr≤L1α_(1n+1n2)(E(‖u0‖2)+∞∑j=1¯‖hj‖2+¯‖g‖2+‖φ1‖2pp−q2pp−q+‖φ2‖2+‖ϕ1‖1)≤L1α_(1n+1n2)(C6+∞∑j=1¯‖hj‖2+¯‖g‖2+‖φ1‖2pp−q2pp−q+‖φ2‖2+‖ϕ1‖1)→0, | (3.30) |
where L1 is the same number of (3.1) and C6>0 is a constant depending only on u0.By φ1∈l2pp−q,φ2∈l2,ϕ1∈l1, (2.6) and (3.1), we infer that
∑‖i‖≥n(¯|gi|2+|φ1,i|2pp−q+|φ2,i|2+|ϕ1,i|)+∑‖i‖≥n∞∑j=1¯|hi,j|2→0 as n→∞. | (3.31) |
It follows from (3.28) to (3.31) that as n→∞,
supu0∈Ksupt≥0(E(‖ξnu(t,0,u0)‖2)+∫t0eα_(r−t)E(∑i∈Zkξ2(‖i‖n)|ui(r,0,u0)|p)dr)→0. | (3.32) |
Then for every ε>0 we can find a number N0=N0(K)∈N satisfying, for all n≥N0 and t≥0,
(E(∑‖i‖≥2n|ui(t,0,u0)|2)+∫t0eα_(r−t)E(∑‖i‖≥2n|ui(t,0,u0)|p)dr)≤(E(‖ξnu(t,0,u0)‖2)+∫t0eα_(r−t)E(∑i∈Zkξ2(‖i‖n)|ui(t,0,u0)|p)dr)≤ε, | (3.33) |
uniformly for u0∈K and t≥0. This concludes the proof.
In the sequel, we use L(u(t,0,u0)) to denote the probability distribution of the solution u(t,0,u0) of (2.15)–(2.16) which has initial condition u0 at initial time 0. Then we have the following tightness of a family of distributions of solutions.
Lemma 4.1. Suppose (2.1)–(2.7) and (3.1) hold. Then the family {L(u(t,0,u0)):t≥0} of the distributions ofthe solutions of (2.15)–(2.16) is tight on l2.
Proof. For simplicity, we will write the solution u(t,0,u0) as u(t) from now on. It follows from Lemma 3.1 that there exists a constant c1>0 such that
E(‖u(t)‖2)≤c1,for allt≥0. | (4.1) |
By Chebyshev's inequality, we get from (4.1) that for all t≥0,
P(‖u(t)‖2≥R)≤c1R2→0asR→∞. |
Hence for every ϵ>0, there exists R1=R1(ϵ)>0 such that for all t≥0,
P{‖u(t)‖2≥R1}≤12ϵ. | (4.2) |
By Lemma 3.2, we infer that for each ϵ>0 and m∈N, there exists an integer nm=nm(ϵ,m) such that for all t≥0,
E(∑|i|>nm|ui(t)|2)<ϵ22m+2, |
and hence for all t≥0 and m∈N,
P({∑|i|>nm|ui(r)|2≥12m})≤2mE(∑|i|>nm|ui(r)|2)<ϵ2m+2. | (4.3) |
It follows from (4.3) for all t≥0,
P(∞∪m=1{∑|i|>nm|ui(t)|2≥12m})≤∞∑m=1ϵ2m+2≤14ϵ, |
which shows that for all t≥0,
P({∑|i|>nm|ui(t)|2≤12mfor allm∈N})>1−ϵ2. | (4.4) |
Given ϵ>0, set
Y1,ϵ={v∈l2:‖v‖≤R1(ϵ)}, | (4.5) |
Y2,ϵ={v∈l2:∑|i|>nm|vi(r)|2≤12mfor allm∈N}, | (4.6) |
and
Yϵ=Y1,ϵ∩Y2,ϵ. | (4.7) |
By (4.2) and (4.4) we get, for all t≥0,
P({u(t)∈Yϵ})>1−ϵ. | (4.8) |
Now, we show the precompactness of {v:v∈Yϵ} in l2. Given κ>0, choose an integer m0=m0(κ)∈N such that 2m0>8κ2. Then by (4.6) we obtain
∑|i|>nm0|vi|2≤12m0<κ28,∀v∈Yϵ. | (4.9) |
On the other hand, by (4.5) we see that the set {(vi)|i|≤m0:v∈Yϵ} is bounded in the finite-dimensional space R2m0+1 and hence precompact. Consequently, {v:v∈Yϵ} has a finite open cover of balls with radius κ2, which along with (4.9) implies that the set {v:v∈Yϵ} has a finite open cover of balls with radius κ in l2. Since κ>0 is arbitrary, we find that the set {v:v∈Yϵ} is precompact in l2. This completes the proof.
If ϕ:l2→R is a bounded Borel function, then for 0≤r≤t and u0∈l2, we set
(pr,tϕ)(u0)=E(ϕ(u(t,r,u0))) |
and
p(r,u0;t,Γ)=(pr,t1Γ)(u0), |
where Γ∈B(l2) and 1Γ is the characteristic function of Γ. The operators ps,t with 0≤s≤t are called the transition operators for the solutions of (2.15)–(2.16). Recall that a probability measure ν on l2 is periodic for (2.15)–(2.16) if
∫l2(p0,t+Tϕ)(u0)dν(u0)=∫l2(p0,tϕ)(u0)dν(u0),∀t≥0. |
Lemma 4.2. [21]Let ϱ(ψ,ω) be a scalar bounded measurable randomfunction of ψ, independent of Fs. Let ς be anFs-measurable random variable. Then
E(ϱ(ς,ω)|Fs)=E(ϱ(ς,ω)). |
The transition operators {pr,t}0≤r≤t have the following properties.
Lemma 4.3. Assume that (2.1)–(2.7) and (3.1) hold. Then:
(i) {pr,t}0≤r≤t is Feller; that is, for every bounded andcontinuous ϕ:l2→R, the function pr,tϕ:l2→Ris also bounded and continuous for all 0≤r≤t.
(ii) The family {pr,t}0≤r≤t is T-periodic; that is, for all 0≤r≤t,
p(r,u0;t,⋅)=p(r+T,u0;t+T,⋅),∀u0∈l2. |
(iii) {u(t,0,u0)}t≥0 is a l2-valued Markov process.
Finally, we present our main result on the existence of periodic measures for problem (2.15)–(2.16).
Theorem 4.4. Assume that (2.1)–(2.7) and (3.1) hold. Then problem (2.15)–(2.16) has a periodic measure on l2.
Proof. We apply Krylov-Bogolyubov's method to prove the existence of periodic measures of (2.15)–(2.16), define a probability measure μn by
μn=1nn∑l=1p(0,0;lT,⋅). | (4.10) |
By Lemma 4.1 we see the sequence {μn}∞n=1 is tight on l2, and hence there exists a probability measure μ on l2 such that, up to a subsequence,
μn→μ,as n→∞. | (4.11) |
By (4.10)–(4.11) and Lemma 4.3, we infer that for every t≥0 and every bounded and continuous function ϕ:l2→R,
∫l2(p0,tϕ)(u0)dμ(u0)=∫l2(∫l2ϕ(y)p(0,u0;t,dy))dμ(u0)=limn→∞1nn∑l=1∫l2(∫l2ϕ(y)p(0,u0;t,dy))p(0,0;lT,du0)=limn→∞1nn∑l=1∫l2(∫l2ϕ(y)p(kT,u0;t+lT,dy))p(0,0;kT,du0)=limn→∞1nn∑l=1∫l2ϕ(y)p(0,0;t+lT,dy)=limn→∞1nn∑l=1∫l2ϕ(y)p(0,0;t+lT+T,dy)=limn→∞1nn∑k=1∫l2(∫l2ϕ(y)p(0,u0;t+T,dy))p(0,0;lT,du0)=∫l2(∫l2ϕ(y)p(0,u0;t+T,dy))dμ(u0)=∫l2(p0,t+Tϕ)(u0)dμ(u0), | (4.12) |
which shows that μ is a periodic measure of (2.15)–(2.16), as desired.
The author declares there is no conflict of interest.
[1] |
J. Brejcha, M. Čadík, State-of-the-art in visual geo-localization, Pattern Anal. Appl., 20 (2017), 613–637. https://doi.org/10.1007/s10044-017-0611-1 doi: 10.1007/s10044-017-0611-1
![]() |
[2] | C. McManus, W. Churchill, W. Maddern, A. D. Stewart, P. Newman, Shady dealings: robust, long-term visual localisation using illumination invariance, in IEEE International Conference on Robotics and Automation (ICRA), (2014), 901–906. https://doi.org/10.1109/icra.2014.6906961 |
[3] | S. Middelberg, T. Sattler, O. Untzelmann, L. Kobbelt, Scalable 6-dof localization on mobile devices, in European Conference on Computer Vision (ECCV), 8690 (2014), 268–283. https://doi.org/10.1007/978-3-319-10605-2_18 |
[4] |
N. Suenderhauf, S. Shirazi, A. Jacobson, F. Dayoub, E. Pepperell, B. Upcroft, et al., Place recognition with ConvNet landmarks: viewpoint-robust, condition-robust, training-free, Rob. Sci. Syst., 2015 (2015), 1–10. https://doi.org/10.15607/RSS.2015.XI.022 doi: 10.15607/RSS.2015.XI.022
![]() |
[5] | J. Hays, A. A. Efros, Im2gps: estimating geographic information from a single image, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2008), 1–8. https://doi.org/10.1109/CVPR.2008.4587784 |
[6] |
A. R. Zamir, M. Shah, Image geo-localization based on multiplenearest neighbor feature matching usinggeneralized graphs, IEEE Trans. Pattern Anal. Mach. Intell., 36 (2014), 1546–1558. https://doi.org/10.1109/TPAMI.2014.2299799 doi: 10.1109/TPAMI.2014.2299799
![]() |
[7] | T. Sattler, M. Havlena, K. Schindler, M. Pollefeys, Large-scale location recognition and the geometric burstiness problem, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2016), 1582–1590. https://doi.org/10.1109/CVPR.2016.175 |
[8] | N. N. Vo, N. Jacobs, J. Hays, Revisiting im2gps in the deep learning era, in IEEE International Conference on Computer Vision (ICCV), (2017), 2621–2630. https://doi.org/10.1109/ICCV.2017.286 |
[9] | M. Noda, T. Takahashi, D. Deguchi, I. Ide, H. Murase, Y. Kojima, et al., Vehicle ego-localization by matching in-vehicle camera images to an aerial image, in Asian Conference on Computer Vision Workshops, 6469 (2010), 163–173. https://doi.org/10.1007/978-3-642-22819-3_17 |
[10] | T. Senlet, A. Elgammal, A framework for global vehicle localization using stereo images and satellite and road maps, in 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), (2011), 2034–2041. https://doi.org/10.1109/ICCVW.2011.6130498 |
[11] | M. Bansal, H. S. Sawhney, H. Cheng, K. Daniilidis, Geo-localization of street views with aerial image databases, in 19th ACM international conference on Multimedia, (2011), 1125–1128. https://doi.org/10.1145/2072298.2071954 |
[12] | T. Y. Lin, S. Belongie, J. Hays, Cross-view image geolocalization, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2013), 891–898. https://doi.org/10.1109/CVPR.2013.120 |
[13] |
D. G. Lowe, Distinctive image features from scale-invariant keypoints, Int. J. Comput. Vision, 60 (2004), 91–110. https://doi.org/10.1023/B:VISI.0000029664.99615.94 doi: 10.1023/B:VISI.0000029664.99615.94
![]() |
[14] | H. Bay, T. Tuytelaars, L. van Gool, Surf: speeded up robust features, in European Conference on Computer Vision, 3951 (2006), 404–417. https://doi.org/10.1007/11744023_32 |
[15] | N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), (2005), 886–893. https://doi.org/10.1109/CVPR.2005.177 |
[16] | E. Shechtman, M. Irani, Matching local selfsimilarities across images and videos, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2007), 1–8. https://doi.org/10.1109/CVPR.2007.383198 |
[17] |
A. Oliva, A. Torralba, Modeling the shape of the scene: a holistic representation of the spatial envelope, Int. J. Comput. Vision, 42 (2017), 145–175. https://doi.org/10.1023/A:1011139631724 doi: 10.1023/A:1011139631724
![]() |
[18] | S. Workman, N. Jacobs, Wide-area image geolocalization with aerial reference imagery, in IEEE International Conference on Computer Vision (ICCV), (2015), 3961–3969. https://doi.org/10.1109/ICCV.2015.451 |
[19] |
A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Commun. ACM, 60 (2017), 84–90. https://doi.org/10.1145/3065386 doi: 10.1145/3065386
![]() |
[20] | S. Hu, M. Feng, R. M. H. Nguyen, G. H. Lee, CVM-Net: cross-view matching network for image-based ground-to-aerial geo-localization, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2018), 7258–7267. https://doi.org/10.1109/CVPR.2018.00758 |
[21] | B. Sun, C. Chen, Y. Zhu, J. Jiang, GEOCAPSNET: ground to aerial view image geo-localization using capsule network, in IEEE International Conference on Multimedia and Expo (ICME), (2019), 742–747. https://doi.org/10.1109/ICME.2019.00133 |
[22] | S. Cai, Y. Guo, S. Khan, J. Hu, G. Wen, Ground-to-aerial image geo-localization with a hard exemplar reweighting triplet loss, in IEEE International Conference on Computer Vision (ICCV), (2019), 8391–8400. https://doi.org/10.1109/ICCV.2019.00848 |
[23] | Y. Shi, X. Yu, L. Liu, T. Zhang, H. Li, Optimal feature transport for cross-view image geo-localization, in AAAI Conference on Artificial Intelligence, 34 (2019), 11990–11997. https://doi.org/10.1609/aaai.v34i07.6875 |
[24] | S. Karen, A. Zisserman, Very deep convolutional networks for large-scale image recognition, preprint, arXiv: 1409.1556. |
[25] | S. Xie, R. Girshick, P. Dollar, Z. Tu, K. He, Aggregated residual transformations for deep neural networks, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2017), 1492–1500. https://doi.org/10.1109/CVPR.2017.634 |
[26] | K. Regmi, M. Shah, Bridging the domain gap for ground-to-aerial image matching, in IEEE International Conference on Computer Vision (ICCV), (2019), 470–479. https://doi.org/10.1109/ICCV.2019.00056 |
[27] | M. Mehdi, S. Osindero, Conditional generative adversarial nets, preprint, arXiv: 1411.1784. |
[28] | Y. Shi, L. Liu, X. Yu, H. Li, Spatial-aware feature aggregation for image based cross-view geo-localization, in 33rd Conference on Neural Information and Processing Systems (NIPS), (2019), 10090–10100. Available from: https://proceedings.neurips.cc/paper/2019/file/ba2f0015122a5955f8b3a50240fb91b2-Paper.pdf. |
[29] | Y. Shi, X. Yu, D. Campbell, H. Li, Where am i looking at? Joint location and orientation estimation by cross-view matching, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2020), 4064–4072. https://doi.org/10.1109/CVPR42600.2020.00412 |
[30] | H. Yang, X. Lu, Y. Zhu, Cross-view geo-localization with layer-to-layer transformer, in 35th Conference on Neural Information and Processing Systems (NIPS), (2021), 29009–29020. Available from: https://papers.nips.cc/paper_files/paper/2021/file/f31b20466ae89669f9741e047487eb37-Paper.pdf. |
[31] |
T. Wang, Z. Zheng, C. Yan, J. Zhang, Y. Sun, B. Zheng, et al., Each part matters: local patterns facilitate cross-view geo-localization, IEEE Trans. Circuits Syst. Video Technol., 32 (2022), 867–879. https://doi.org/10.1109/TCSVT.2021.3061265 doi: 10.1109/TCSVT.2021.3061265
![]() |
[32] | A. Toker, Q. Zhou, M. Maximov, L. Leal-Taixé, Coming down to earth: satellite-to-street view synthesis for geo-localization, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2021), 6484–6493. https://doi.org/10.1109/CVPR46437.2021.00642 |
[33] | M. Tan, Q. Le, Efficientnetv2: smaller models and faster training, in 38th International Conference on Machine Learning (ICML), preprint, arXiv: 2104.00298. |
[34] | D. Misra, T. Nalamada, A. U. Arasanipalai, Q. Hou, Rotate to attend: convolutional triplet attention module, in IEEE/CVF Winter Conference on Applications of Computer Vision, (2021), 3139–3148. https://doi.org/10.1109/WACV48630.2021.00318 |
[35] | N. N. Vo, J. Hays, Localizing and orienting street views using overhead imagery, in European Conference on Computer Vision (ECCV), 9905 (2016), 494–509. https://doi.org/10.1007/978-3-319-46448-0_30 |
[36] | L. Liu, H. Li, Lending orientation to neural networks for cross-view geo-localization, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2019), 5624–5633. Available from: https://openaccess.thecvf.com/content_CVPR_2019/papers/Liu_Lending_Orientation_to_Neural_Networks_for_Cross-View_Geo-Localization_CVPR_2019_paper.pdf. |
[37] | R. Rodrigues, M. Tani, Are these from the same place? Seeing the unseen in cross-view image geo-localization, in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), (2021), 3753–3761. https://doi.org/10.1109/WACV48630.2021.00380 |
[38] | A. Hermans, L. Beyer, B. Leibe, In defense of the triplet loss for person re-identification, preprint, arXiv: 1703.07737. |
[39] | M. Tan, Q. Le, Efficientnet: rethinking model scaling for convolutional neural networks, International conference on machine learning, preprint, arXiv: 1905.11946. |
[40] | B. Zoph, Q. V. Le, Neural architecture search with reinforcement learning, preprint, arXiv: 1611.01578. |
[41] | M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, et al., MnasNet: platform-aware neural architecture search for mobile, in IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2019), 2820–2828. https://doi.org/10.1109/CVPR.2019.00293 |
[42] | S. Gupta, M. Tan, Efficientnet-edgetpu: creating accelerator-optimized neural networks with automl, Google AI Blog, 2019. Available from: https://torontoai.org/2019/08/05/efficientnet-edgetpu-creating-accelerator-optimized-neural-networks-with-automl/. |
[43] | M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L. Chen, Mobilenetv2: inverted residuals and linear bottlenecks, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2018), 4510–4520. https://doi.org/10.1109/CVPR.2018.00474 |
[44] |
J. Hu, L. Shen, S. Albanie, G. Sun, E. Wu, Squeeze-and excitation networks, IEEE Trans. Pattern Anal. Mach. Intell., 42 (2020), 2011–2023. https://doi.org/10.1109/TPAMI.2019.2913372 doi: 10.1109/TPAMI.2019.2913372
![]() |
[45] | J. Park, S. Woo, J. Y. Lee, I. S. Kweon, BAM: bottleneck attention module, preprint, arXiv: 1807.06514. |
[46] | S. Woo, J. Park, J. Y. Lee, I. S. Kweon, CBAM: convolutional block attention module, in European Conference on Computer Vision (ECCV), 11211 (2018), 3–19. https://doi.org/10.1007/978-3-030-01234-2_1 |
[47] | M. Zhai, Z. Bessinger, S. Workman, N. Jacobs, Predicting ground-level scene layout from aerial imagery, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (2017), 867–875. https://doi.org/10.1109/CVPR.2017.440 |
1. | Xintao Li, Rongrui Lin, Lianbing She, Periodic measures for a neural field lattice model with state dependent superlinear noise, 2024, 32, 2688-1594, 4011, 10.3934/era.2024180 | |
2. | Xintao Li, Lianbing She, Jingjing Yao, Periodic measures of fractional stochastic discrete wave equations with nonlinear noise, 2024, 57, 2391-4661, 10.1515/dema-2024-0078 | |
3. | Hailang Bai, Mingkai Yuan, Dexin Li, Yunshun Wu, Weak and Wasserstein convergence of periodic measures of stochastic neural field lattice models with Heaviside ’s operators and locally Lipschitz Lévy noises, 2025, 143, 10075704, 108602, 10.1016/j.cnsns.2025.108602 |