prompt
stringlengths
0
256
answer
stringlengths
1
256
section with the following estimate. \begin{lemma} \label{lem:M3vsM1} Under the assumptions and notation of Theorem \ref{thm:genus_2constr}, \[ M_{\alpha,3} \leq 2^{-(p-1)/2} M_{\alpha,1}. \] \end{lemma} \begin{proof} Let $A\in \Sigma^{(3)}_{\alpha}$, acco
rding to Definition \ref{def:genus}. Notice that the map \[ A\ni u \mapsto \left(\int_\Omega |u|u , \int_\Omega |u|^p u\right)\in{\mathbb{R}}^2 \] is continuous and equivariant. By the Borsuk-Ulam Theorem, we deduce the existence of $u_a\in A$ such that \[
\int_\Omega |u^+_a|^2 = \int_\Omega |u^-_a|^2 = \frac12,\qquad \int_\Omega |u^+_a|^{p+1} = \int_\Omega |u^-_a|^{p+1} = \frac12 \int_\Omega |u_a|^{p+1}, \] while \[ \text{either }\int_\Omega |\nabla u^+_a|^2 \leq \frac{\alpha}{2} \qquad \text{or }\int_\Ome
ga |\nabla u^-_a|^2 \leq \frac{\alpha}{2}. \] For concreteness let us assume that the first alternative holds; as a consequence, we obtain that $v:=\sqrt2 u_a^+$ belongs to $\overline{\mathcal{B}}_\alpha$. This yields \[ M_{\alpha,1}\geq \int_\Omega |v|^{p
+1} = 2^{(p+1)/2} \int_\Omega |u_a^+|^{p+1} = \frac{2^{(p+1)/2}}{2} \int_\Omega |u_a|^{p+1} \geq 2^{(p-1)/2} \inf_{u\in A} \int_\Omega |u|^{p+1}, \] and since $A\in \Sigma^{(3)}_{\alpha}$ is arbitrary the proposition follows. \end{proof} \section{Min-m
ax principles on the unit sphere in \texorpdfstring{$L^2$}{L\texttwosuperior}}\label{sec:1const} According to equation \eqref{Emu}, let ${\mathcal{M}}\subset H^1_0(\Omega)$ denote the unit sphere with respect to the $L^2$ norm and $\mathcal{E}_{\mu}$ the
energy functional associated to \eqref{eq:main_prob_u}. In this section we are concerned with critical points of $\mathcal{E}_{\mu}$ on ${\mathcal{M}}$ (which, in turn, correspond to solutions of our starting problem \eqref{eq:main_prob_U}). By the Gaglia
rdo-Nirenberg inequality \eqref{sobest}, setting $\|\nabla u\|^2_{L^2}=\alpha$, one obtains \begin{equation}\label{eq:boundonboundEmu} \frac12\,\alpha- \mu\frac{C_{N,p}}{p+1}\,\alpha^{N(p-1)/4} \leq \mathcal{E}_{\mu}(u)\le \frac12\alpha. \end{equation} In
particular, $\mathcal{E}_{\mu}$ is bounded on any bounded subset of ${\mathcal{M}}$, and it is bounded from below (and coercive) on the entire ${\mathcal{M}}$ for {subcritical} $p<1+4/N$ and for {critical} $p=1+4/N$ whenever $\mu< \frac{p+1}{2}C_{N,p}^{-1}
$ . In these cases, one can easily show that $\mathcal{E}_{\mu}$ satifies the P.S. condition and apply the classical {minimax principle for even functionals} on a closed symmetric submanifold (see e.g. \cite[Thm. II.5.7]{St_2008}). In the complementary ca
se, when $p$ is either supercritical, i.e. $p>1+4/N$, or critical and $\mu$ is large, then $\mathcal{E}_{\mu}$ is not bounded from below (see e.g. \eqref{minusinfty} below). In order to provide a minimax principle suitable for this case, we recall the Defi
nition \ref{def:genus} of genus and that of $\mathcal{B}_\alpha$ (see equation \eqref{eq:defBU}). Furthermore, we denote with $K_{c}$ the (closed and symmetric) set of critical points of ${\mathcal{E}}_\mu$ at level $c$ contained in $\mathcal{B}_\alpha$. T
he following theorem is an adaptation of well known arguments of previous critical point theorems relying on index theory. \begin{theorem} \label{infsupteo} Let $k\ge1$, $\alpha>\lambda_k(\Omega)$, $\mu>0$ and $\tau>0$ be fixed, and let $c_k$ be defined as
in Theorem \ref{thm:genus_1constr}, equation \eqref{infsuplev}. If \begin{equation} \label{ass2} c_k < \hat c_k:= \inf_{\substack{A\in\Sigma^{(k)}_{\alpha}\\ A\setminus\mathcal{B}_{\alpha-\tau}\neq\emptyset }} \sup_{A\setminus\mathcal{B}_{\alpha-\tau}}{\
mathcal{E}}_\mu, \end{equation} then $K_{c_k}\neq\emptyset$, and it contains a critical point of Morse index less or equal to $k$. \end{theorem} \begin{remark} In case assumption \eqref{ass2} holds for $k,k+1,\dots,k+r$, and $c=c_k=...c_{k+r}$, then it is
standard to extend Theorem \ref{infsupteo} to obtain \begin{equation} \label{indexK} \gamma(K_c)\ge r+1, \end{equation} so that $K_c$ contains infinitely many critical points. \end{remark} \begin{proof}[Proof of Theorem \ref{infsupteo}] For any $a\in{\math
bb{R}}$ we denote by ${\mathcal{M}}_a$ the sublevel set $\{\mathcal{E}_\mu<a\}$. First of all we notice that both $c_k$ and $\hat c_k$ are well defined and finite, by Lemma \ref{lemma:genusbigger} and equation \eqref{eq:boundonboundEmu}. Suppose now by con
tradiction that $K_{c_k}=\emptyset$. By a suitably modified version of the Deformation Lemma (recall that ${\mathcal{E}}_\mu$ satisfies the P.S. condition on ${\mathcal{M}}$), there exist $\delta>0$ and an equivariant homeomorphism $\eta$ such that $\eta(u
)=u$ outside $\mathcal{B}_{\alpha}\cap {\mathcal{M}}_{c_k+2\delta}$ and \begin{equation} \label{lowlev} \eta({{\mathcal{M}}_{c_k+\delta}\cap \mathcal{B}_{\alpha-\tau}})\subset {\mathcal{M}}_{c_k-\delta}\cap \mathcal{B}_{\alpha}. \end{equation} By definitio
n of $c_{k}$ there exists $A\in \Sigma^{(k)}_{\alpha}$ such that $A\subset {\mathcal{M}}_{c_k+\delta}$; it follows by assumption \eqref{ass2} (and by decreasing $ \delta$ if necessary) that $A\subset {\mathcal{M}}_{c_k+\delta}\cap \mathcal{B}_{\alpha-\tau}
$. Then, since $\eta$ is an odd homeomorphism, $\eta(A)\in \Sigma^{(k)}_{\alpha}$ and, by definition, $\sup_{\eta(A)}{\mathcal{E}}_\mu \ge c_k$, in contradiction with \eqref{lowlev}. Finally, the estimate of the Morse index is a direct consequence of the d
efinition of genus we deal with: see \cite{MR968487}, Proposition on page 1030, or the discussion at the end of Section 2 in \cite{MR991264}. \end{proof} We now provide a sufficient condition to guarantee the validity of assumption \eqref{ass2}. \begin{lem
ma}\label{lem:ckMak} Let $k\ge1$, $\alpha>\lambda_k(\Omega)$ and $\mu>0$ satisfy \begin{equation} \label{muboundef} 0<\mu<\frac{p+1}{2}\,\frac{\alpha-\lambda_{k}(\Omega)}{M_{\alpha,k}-|\Omega|^{-\frac{p-1}{2}}} \end{equation} where $M_{\alpha,k}$ is define
d in Theorem \ref{thm:genus_2constr}. Then, for $\tau>0$ sufficiently small, \eqref{ass2} holds true. \end{lemma} \begin{proof} We first estimate $c_k$ from above. To this aim, we construct a subset $\tilde A\in \Sigma^{(k)}_{\alpha-\tau}$ (for any $\tau$
sufficiently small) as \begin{equation} \label{Atilde} \tilde A= \left\{\sum_{i=1}^k x_i\varphi_i : x=(x_1,\dots,x_k)\in{\mathbb{S}}^{k-1}\right\}, \end{equation} where, as usual $\varphi_i$ denotes the Dirichlet eigenfunction associated to $\lambda_i(\Ome
ga)$. Indeed $\gamma(\tilde A)=k$ (it is homeomorphic to ${\mathbb{S}}^{k-1}$), and $\max_{u\in\tilde A}\|u\|^2_{H^1_0}=\lambda_{k}(\Omega)<\alpha-\tau$ for $\tau$ small. Hence Holder inequality yields \begin{equation} \label{boundabove1} c_k \leq \sup_{\t
ilde A}\mathcal{E}_{\mu}\le \frac{1}{2}\lambda_{k}(\Omega) - \frac{\mu}{p+1}\,|\Omega|^{-\frac{p-1}{2}}. \end{equation} On the other hand, let $A\in\Sigma^{(k)}_{\alpha}$. Theorem \ref{thm:genus_2constr} implies \[ \inf_{u\in A} \int_\Omega |u|^{p+1} \leq
M_{\alpha,k}. \] If moreover $A\setminus\mathcal{B}_{\alpha-\tau}\neq\emptyset$ we infer \[ \sup_{A\setminus\mathcal{B}_{\alpha-\tau}}\mathcal{E}_{\mu}\ge \frac{1}{2}(\alpha - \tau) - \frac{\mu}{p+1} M_{\alpha,k}, \] and taking the infimum an analogous in
equality holds true for $\hat c_k$. Comparing with \eqref{boundabove1} the lemma follows. \end{proof} Exploiting the results above, we are ready to prove our main existence results. \begin{proof}[End of the proof of Theorem \ref{thm:genus_1constr}] By Theo
rem \ref{infsupteo} and Lemma \ref{lem:ckMak} the proof is completed by choosing \[ \hat\mu_k:=\sup_{\alpha>\lambda_k(\Omega)} \frac{p+1}{2}\,\frac{\alpha-\lambda_{k}(\Omega)}{M_{\alpha,k}-|\Omega|^{-\frac{p-1}{2}}}. \qedhere \] \end{proof} \begin{proof}[P
roof of Theorem \ref{thm:intro_GS}] We write the proof in terms of ${\mathcal{E}}_\mu$, the theorem following by the relations in \eqref{eq:main_prob_u}. Recall that, for every $u\in \overline{\mathcal{B}}_\alpha$, $\gamma \left(\{u,-u\}\right)=1$. We dedu
ce that $c_1$ is actually a local minimum for ${\mathcal{E}}_\mu$, achieved by some $u$ which solves \eqref{eq:main_prob_u} (for a suitable $\lambda$), and it can be chosen positive by symmetry. Since \[ \int_\Omega |\nabla u|^2 + \lambda u^2 - p\mu|u|^{p+
1}\,dx=-(p-1)\int_\Omega \mu|u|^{p+1}\,dx<0, \] and $H^1_0(\Omega) = \spann\{u\}\oplus T_{u}\mathcal{M}$, we have that $u$ has Morse index $1$. In a standard way, the minimality property of $u$ implies also orbital stability of the associated solitary wav
e (see e.g. \cite{MR677997}). Turning to the estimates for $\hat\mu_1 = \hat\rho_1^{(p-1)/2}$, we can deduce it using Lemma \ref{lem:ckMak} and Remark \ref{rem:MvsCNp}, which yield \[ \hat\mu_1\left(\Omega,p\right):=\sup_{\alpha>\lambda_1(\Omega)} \frac{p+
1}{2}\,\frac{\alpha-\lambda_{1}(\Omega)}{C_{N,p} \alpha^{\frac{N(p-1)}{4}}-|\Omega|^{-\frac{p-1}{2}}} \geq \frac{p+1}{2C_{N,p}}\,\sup_{\alpha>\lambda_1(\Omega)}\frac{\alpha-\lambda_{1} (\Omega)}{\alpha^{\beta}}, \] where $\beta:=N(p-1)/4$. Now, if $\beta\l
eq1$ we obtain the desired bound for the subcritical and critical cases. On the other hand, when $\beta>1$, elementary calculations show that \[ \hat\mu_1\left(\Omega,p\right)\geq \frac{p+1}{2C_{N,p}} \,\frac{(\beta-1)^{(\beta-1)}}{\beta^\beta}\, \lambda_1
(\Omega)^{-(\beta-1)}, \] and finally \[ \hat\rho_1\left(\Omega,p\right)\geq \underbrace{\left[\frac{p+1}{2C_{N,p}} \,\frac{(\beta-1)^{(\beta-1)}}{\beta^\beta}\right]^{\frac{2}{p-1}}}_{D_{N,p}}\, \lambda_1(\Omega)^{\frac{2}{p-1}-\frac{N}{2}}.\qedhere \] \e
nd{proof} \begin{proof}[Proof of Proposition \ref{thm:intro_3>1}] As usual, by \eqref{eq:main_prob_u}, we have to prove that \[ \hat\mu_3\left(\Omega,p\right)\geq 2^{(p-1)/2} D_{N,p}\lambda_3(\Omega)^{\frac{2}{p-1}-\frac{N}{2}}. \] By Lemmas \ref{lem:ckMak
}, \ref{lem:M3vsM1}, and Remark \ref{rem:MvsCNp} we obtain \[ \begin{split} \hat\mu_3 &= \sup_{\alpha>\lambda_3(\Omega)} \frac{p+1}{2}\,\frac{\alpha-\lambda_{3}(\Omega)}{M_{\alpha,3}-|\Omega|^{-\frac{p-1}{2}}} \geq \sup_{\alpha>\lambda_3(\Omega)} \frac{p+
1}{2}\,\frac{\alpha-\lambda_{3}(\Omega)}{ 2^{-(p-1)/2}M_{\alpha,1}-|\Omega|^{-\frac{p-1}{2}}}\\ &\geq 2^{(p-1)/2}\sup_{\alpha>\lambda_3(\Omega)} \frac{p+1}{2}\,\frac{\alpha-\lambda_{3} (\Omega)}{C_{N,p}\alpha^\beta-2^{(p-1)/2}|\Omega|^{-\frac{p-1}{2}}}, \e
nd{split} \] where $\beta:=N(p-1)/4$, and the desired result follows by arguing as in the proof of Theorem \ref{thm:intro_GS}. \end{proof} To conclude this section we prove that in the supercritical case, if $\mu$ is not too large, in addition to $(c_k)_k$
there is a further sequence of critical levels $(\bar c_k)$ of $\mathcal{E}_{\mu}$ constrained to $\mathcal{M}$. For concreteness, let us first consider the case $k=1$: since in such case $c_1$ is a local minimum of ${\mathcal{E}}_\mu$ in $\mathcal{M}$, a
nd ${\mathcal{E}}_\mu$ is unbounded from below in $\mathcal{M}$, the critical level $\bar c_1$ is of mountain pass type. \begin{proposition} \label{mpcritlev} Let $p>1+4/N$, $\mu<\hat\mu_1$, and $u_1$ denote the local minimum point of ${\mathcal{E}}_\mu$ i
n $\mathcal{M}$, according to Theorems \ref{infsupteo} and \ref{thm:intro_GS}. The value \[ \bar c_1 : =\inf_{\gamma\in \Gamma}\sup_{[0,1]}\mathcal{E}_{\mu}(\gamma(s)), \quad\text{where }\Gamma:=\left\{\gamma\in C([0,1];{\mathcal{M}}) : \gamma(0)=u_1,\,\ga
mma(1)<c_1-1\right\}, \] is a critical level for ${\mathcal{E}}_\mu$ in $\mathcal{M}$. \end{proposition} \begin{proof} Notice that, if $p>1+4/N$, then $\mathcal{E}_{\mu}\to -\infty$ along some sequence in ${\mathcal{M}}$. Indeed, by defining \begin{equati
on} \label{concfun} w_n(x): = \eta(x)Z_{N,p}\big ((x-x_0)/a_n\big )\quad \text{and}\quad \tilde{w_n}:=\frac{w_n}{\| w_n\|^2_{L^2(\Omega)}}\,\in\, {\mathcal{M}}, \end{equation} where $a_n\to 0^+$, $x_0\in \Omega$ and $\eta\in \mathcal{C}_0^{\infty}(\Omega)$
, $\eta(x_0)=1$, we obtain \begin{equation}\label{minusinfty} \alpha_n :=\|\nabla \tilde w_n\|^2_{L^2(\Omega)}\to +\infty, \qquad \frac{\int_{\Omega} |\tilde w_n|^{p+1} \,dx}{\alpha_n^{N(p-1)/4}}\to C_{N,p}, \qquad \mathcal{E}_{\mu}(\tilde w_n)\to -\infty
\end{equation} for $n\to +\infty$. Since $u_1$ is a local minimum, the functional $\mathcal{E}_{\mu}$ has a mountain pass structure on ${\mathcal{M}}$; by recalling that $\mathcal{E}_{\mu}$ satisfies the P.S. condition the proposition follows. \end{proof}
\begin{remark}\label{rem:further_crit_lev} One can generalize Proposition \ref{mpcritlev} by constructing critical points via a saddle-point theorem in the following way: let us pick $k$ points $x_1, x_2,...,x_k$ in $\Omega$ and consider the corresponding
function $\tilde w_i$; we may assume that $\mathrm{supp}\,\tilde w_i\cap \mathrm{supp}\,\tilde w_j=\emptyset$ for $i\neq j$, so that these functions are orthogonal. Let us now define the subspace $V_k=\mathrm{span}\{\varphi_1,\dots,\varphi_k;\tilde w_1,...
,\tilde w_k\}$; note that dim $V_k=2k$. Let $R$ be an operator (in $L^2(\Omega)$) such that $R=I$ on $V_k^{\perp}$, $Ru_i=\tilde w_i$, $i=1,2,..,k$. Possibly after permutations, we can choose $R$ such that $R\big |_{V_k}\in SO(2k)$ (actually, there are inf
initely many different choices of $R$). Now, since $SO(2k)$ is (arcwise) connected, there is a continuous path $\tilde{\gamma}:\,[0,1]\rightarrow SO(2k)$ such that $\gamma(0)=I$, $\gamma(1)=R\big |_{V_k}$. Then, we can define the following map \[ \gamma:\,
[0,1] \times S^{k-1}\rightarrow {\mathcal{M}}, \quad\quad \gamma(s;t_1,....,t_k)=\sum_{i=1}^{k}t_i\tilde{\gamma}(s)u_i,\quad \] where $\sum_{i=1}^{k}t_i^2=1$. It is clear that $\gamma$ is continuous; moreover, \[ \gamma(0;t_1,....,t_k)\in \mathrm{span }\{
\varphi_1,\dots,\varphi_k\}\cap {\mathcal{M}} \quad \mathrm{and}\quad \gamma(1;t_1,....,t_k)\in \mathrm{span }\{\tilde w_1,\dots,\tilde w_k\}\cap {\mathcal{M}}. \] Then, by denoting with $\Gamma_k$ the set of the above paths, if $\mu$ is sufficiently small
we obtain the critical levels \[ \bar c_k : =\inf_{\gamma\in \Gamma_k}\sup_{[0,1]\times S^{k-1}}\mathcal{E}_{\mu}(\gamma(s;t_1,....,t_k)). \] \end{remark} \section{Results in symmetric domains}\label{sec:symm} This section is devoted to the proof of T
heorem \ref{pro:symm}, therefore we assume $1+4/N \leq p < 2^*-1$. We perform the proof in the case of $\Omega=B$, but it will be clear that the main assumption on $\Omega$ is the following: \begin{itemize} \item[\textbf{(T)}] there is a tiling of $\Omega
$, made by $h$ copies of a subdomain $D$, in such a way that from any solution $U_D$ of \eqref{eq:main_prob_U} on $D$ one can construct, using reflections, a solution $U_\Omega$ of \eqref{eq:main_prob_U} on $\Omega$. \end{itemize} Then $U_\Omega$ has $h$
times the mass of $U_D$, and recalling Theorem \ref{thm:intro_GS} we deduce that \eqref{eq:main_prob_U} on $\Omega$ is solvable for any $\rho< h \cdot D_{N,p} \lambda_1(D)^{\frac{2}{p-1}-\frac{N}{2}}$. At this point, for a sequence $(D_k,h_k)_k$ of tiling
s satisfying \textbf{(T)}, we obtain the solvability of \eqref{eq:main_prob_U} on $\Omega$ whenever \[ \rho< h_k \cdot D_{N,p} \lambda_1(D_k)^{\frac{2}{p-1}-\frac{N}{2}}, \] and if we can show that \begin{equation}\label{eq:finaltarget} \frac{ h_k }{ \lam
bda_1(D_k)^{\frac{N}{2}-\frac{2}{p-1}}} \to +\infty\qquad\text{as }k\to+\infty, \end{equation} we deduce the solvability of \eqref{eq:main_prob_U} on $\Omega$ for every mass. Having this scheme in mind, it is easy to prove analogous results on rectangles a
nd also in other kind of domains. Then let $B\subset{\mathbb{R}}^N$ be the ball (w.l.o.g. of radius one), and let \[ D_k:=\left\{(r\cos\theta, r\sin\theta,x_3,\dots,x_N)\in B: - \frac{\pi}{k} < \theta < \frac{\pi}{k}\right\} \] Then $D_k$ satisfies \text
bf{(T)}, with $h_k=k$. In order to estimate $\lambda_1(D_k)$ we observe that, by elementary trigonometry, \[ B'_k = B_{\frac{\sin(\pi/k)}{\sin(\pi/k)+1}}\left(\frac{1}{\sin(\pi/k)+1},0,0,\dots,0\right) \subset D_k, \] and therefore \[ \lambda_1(D_k) \le \l
ambda_1(B'_k) \le C k^2, \] for some dimensional constant $C=C(N)$ and $k$ large. Then \[ \frac{ h_k }{ \lambda_1(D_k)^{\frac{N}{2}-\frac{2}{p-1}}}\ge C \frac{k}{k^{{N}-\frac{4}{p-1}}} = C k^{1-{N}+\frac{4}{p-1}} = C k^{\frac{N-1}{p-1}\left[1+\frac{4}{N-1
} - p\right]}, \] and finally \eqref{eq:finaltarget} holds true whenever $p< 1+\frac{4}{N-1}$, thus completing the proof of Theorem \ref{pro:symm}. \small \subsection*{Acknowledgments} We would like to thank Jacopo Bellazzini, who pointed out that the
results in \cite{MR3318740}, in the supercritical case, can be read in terms of a local minimization. We would also like to thank Benedetta Noris, who read a preliminary version of this manuscript. This work is partially supported by the PRIN-2
012-74FYK7 Grant: ``Variational and perturbative aspects of nonlinear differential problems'', by the ERC Advanced Grant 2013 n. 339958: ``Complex Patterns for Strongly Interacting Dynamical Systems - COMPAT'', and by the INDAM-GNAMPA group.
\section{Introduction} \label{sec:intro} Despite the immense popularity and availability of online video content via outlets such as Youtube and Facebook, most work on object detection focuses on static images. Given the breakthroughs of deep convolution
al neural networks for detecting objects in static images, the application of these methods to video might seem straightforward. However, motion blur and compression artifacts cause substantial frame-to-frame variability, even in videos that appear smo
oth to the eye. These attributes complicate prediction tasks like classification and localization. Object-detection models trained on images tend not to perform competitively on videos owing to domain shift factors \cite{KalogeitonFS15}. Moreover, obje
ct-level annotations in popular video data-sets can be extremely sparse, impeding the development of better video-based object detection models. Girshik \emph{et al}\bmvaOneDot \cite{RCNN_girshick14CVPR} demonstrate that even given scarce labeled tra
ining data, high-capacity convolutional neural networks can achieve state of the art detection performance if first pre-trained on a related task with abundant training data, such as 1000-way ImageNet classification. Followed the pretraining, the netwo
rks can be fine-tuned to a related but distinct domain. Also relevant to our work, the recently introduced models Faster R-CNN \cite{Faster_RCNN_RenHG015} and You Look Only Once (YOLO) \cite{YOLO_RedmonDGF15} unify the tasks of classification and localiz
ation. These methods, which are accurate and efficient, propose to solve both tasks through a single model, bypassing the separate object proposal methods used by R-CNN \cite{RCNN_girshick14CVPR}. In this paper, we introduce a method to extend unified o
bject recognition and localization to the video domain. Our approach applies transfer learning from the image domain to video frames. Additionally, we present a novel recurrent neural network (RNN) method that refines predictions by exploiting contextual
information in neighboring frames. In summary, we contribute the following: \begin{itemize} \item A new method for refining a video-based object detection consisting of two parts: (i) a \emph{pseudo-labeler}, which assigns provisional labels to all avai
lable video frames. (ii) A recurrent neural network, which reads in a sequence of provisionally labeled frames, using the contextual information to output refined predictions. \item An effective training strategy utilizing (i) category-level weak-supervi
sion at every time-step, (ii) localization-level strong supervision at final time-step (iii) a penalty encouraging prediction smoothness at consecutive time-steps, and (iv) similarity constraints between \emph{pseudo-labels} and prediction output at every
time-step. \item An extensive empirical investigation demonstrating that on the YouTube Objects \cite{youtube-Objects} dataset, our framework achieves mean average precision (mAP) of $68.73$ on test data, compared to a best published result of $37.41$ \c
ite{Tripathi_WACV16} and $61.66$ for a domain adapted YOLO network \cite{YOLO_RedmonDGF15}. \end{itemize} \section{Methods} \label{sec:method} In this work, we aim to refine object detection in video by utilizing contextual information from nei
ghboring video frames. We accomplish this through a two-stage process. First, we train a \emph{pseudo-labeler}, that is, a domain-adapted convolutional neural network for object detection, trained individually on the labeled video frames. Specifically,
we fine-tune the YOLO object detection network \cite{YOLO_RedmonDGF15}, which was originally trained for the 20-class PASCAL VOC \cite{PASCAL_VOC} dataset, to the Youtube-Video \cite{youtube-Objects} dataset. When fine-tuning to the 10 sub-categories
present in the video dataset, our objective is to minimize the weighted squared detection loss (equation \ref{eqn:obj_det_loss}) as specified in YOLO \cite{YOLO_RedmonDGF15}. While fine-tuning, we learn only the parameters of the top-most fully-connec
ted layers, keeping the $24$ convolutional layers and $4$ max-pooling layers unchanged. The training takes roughly 50 epochs to converge, using the RMSProp \cite{RMSProp} optimizer with momentum of $0.9$ and a mini-batch size of $128$. As with YOLO \ci
te{YOLO_RedmonDGF15}, our fine-tuned $pseudo-labeler$ takes $448 \times 448$ frames as input and regresses on category types and locations of possible objects at each one of $S \times S$ non-overlapping grid cells. For each grid cell, the model output
s class conditional probabilities as well as $B$ bounding boxes and their associated confidence scores. As in YOLO, we consider a \emph{responsible} bounding box for a grid cell to be the one among the $B$ boxes for which the predicted area and the grou
nd truth area shares the maximum Intersection Over Union. During training, we simultaneously optimize classification and localization error (equation \ref{eqn:obj_det_loss}). For each grid cell, we minimize the localization error for the \emph{respon
sible} bounding box with respect to the ground truth only when an object appears in that cell. Next, we train a Recurrent Neural Network (RNN), with Gated Recurrent Units (GRUs) \cite{Cho14_GRU}. This net takes as input sequences of \emph{pseudo-labe
ls}, optimizing an objective that encourages both accuracy on the target frame and consistency across consecutive frames. Given a series of \emph{pseudo-labels} $\mathbf{x}^{(1)}, ..., \mathbf{x}^{(T)}$, we train the RNN to generate improved predicti
ons $\hat{\mathbf{y}}^{(1)}, ..., \hat{\mathbf{y}}^{(T)}$ with respect to the ground truth $\mathbf{y}^{(T)}$ available only at the final step in each sequence. Here, $t$ indexes sequence steps and $T$ denotes the length of the sequence. As output, we
use a fully-connected layer with a linear activation function, as our problem is regression. In our final experiments, we use a $2$-layer GRU with $150$ nodes per layer, hyper-parameters determined on validation data. The following equations defi
ne the forward pass through a GRU layer, where $\mathbf{h}^{(t)}_l$ denotes the layer's output at the current time step, and $\mathbf{h}^{(t)}_{l-1}$ denotes the previous layer's output at the same sequence step: \begin{equation} \label{eqn:GRU} \begin{al
igned} \mathbf{r}^{(t)}_l &= \sigma(\mathbf{h}^{(t)}_{l-1}W^{xr}_l + \mathbf{h}^{(t-1)}_lW^{hr}_l + \mathbf{b}^r_l)\\ \mathbf{u}^{(t)}_l &= \sigma(\mathbf{h}^{(t)}_{l-1}W^{xu}_l + \mathbf{h}^{(t-1)}_lW^{hu}_l + \mathbf{b}^u_l)\\ \mathbf{c}^{(t)}_l &= \sigm
a(\mathbf{h}^{(t)}_{l-1}W^{xc}_l + r_t \odot(\mathbf{h}^{(t-1)}_lW^{hc}_l) + \mathbf{b}^c_l)\\ \mathbf{h}^{(t)}_l &= (1-\mathbf{u}^{(t)}_l)\odot \mathbf{h}^{(t-1)}_l + \mathbf{u}^{(t)}_l\odot \mathbf{c}^{(t)}_l \end{aligned} \end{equation} Here, $\sigma$ d
enotes an element-wise logistic function and $\odot$ is the (element-wise) Hadamard product. The reset gate, update gate, and candidate hidden state are denoted by $\textbf{r}$, $\textbf{u}$, and $\textbf{c}$ respectively. For $S = 7$ and $B=2$, the pseu
do-labels $\mathbf{x}^{(t)}$ and prediction $\hat{\mathbf{y}}^{(t)}$ both lie in $\mathbb{R}^{1470}$. \vspace{-2.5mm} \subsection{Training} We design an objective function (Equation \ref{eqn:objective}) that accounts for both accuracy at the target frame
and consistency of predictions across adjacent time steps in the following ways: \begin{equation} \label{eqn:objective} \mbox{loss} = \mbox{d\_loss} + \alpha \cdot \mbox{s\_loss} + \beta \cdot \mbox{c\_loss} + \gamma \cdot \mbox{pc\_loss} \end{equation
} Here, d\_loss, s\_loss, c\_loss and pc\_loss stand for detection\_loss, similarity\_loss, category\_loss and prediction\_consistency\_loss described in the following sections. The values of the hyper-parameters $\alpha=0.2$, $\beta=0.2$ and $\gamma=0.1$
are chosen based on the detection performance on the validation set. The training converges in 80 epochs for parameter updates using RMSProp \cite{RMSProp} and momentum $0.9$. During training we use a mini-batch size of $128$ and sequences of length
$30$. \subsubsection{Strong Supervision at Target Frame} On the final output, for which the ground truth classification and localization is available, we apply a multi-part object detection loss as described in YOLO \cite{YOLO_RedmonDGF15}. \vspace{-2.5
mm} \begin{equation} \label{eqn:obj_det_loss} \begin{aligned} \mbox{detection\_loss} &= \lambda_{coord}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj}_{ij}\big(\mathit{x}^{(T)}_i - \hat{\mathit{x}}^{(T)}_i\big)^2 + \big(\mathit{y}^{(T)}_i - \hat{\mathit{y}
}^{(T)}_i\big)^2 \\ & + \lambda_{coord}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj}_{ij}\big(\sqrt{w_i}^{(T)} - \sqrt{\hat{w}^{(T)}_i}\big)^2 + \big (\sqrt{h_i}^{(T)} - \sqrt{\hat{h}^{(T)}_i} \big)^2 \\ & + \sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{obj
}_{ij}(\mathit{C}_i - \hat{\mathit{C}_i})^2 \\ & + \lambda_{noobj}\sum^{S^2}_{i=0}\sum^{B}_{j=0}\mathbbm{1}^{noobj}_{ij}\big(\mathit{C}^{(T)}_i - \hat{\mathit{C}}^{(T)}_i\big)^2 \\ & + \sum^{S^2}_{i=0}\mathbbm{1}^{obj}_{i}\sum_{c \in classes}\big(p_i^{(T)}
(c) - \hat{p_i}^{(T)}(c)\big)^2 \end{aligned} \end{equation} where $\mathbbm{1}^{obj}_{i}$ denotes if the object appears in cell $i$ and $\mathbbm{1}^{obj}_{ij}$ denotes that $j$th bounding box predictor in cell $i$ is \emph{responsible} for that pred
iction. The loss function penalizes classification and localization error differently based on presence or absence of an object in that grid cell. $x_i, y_i, w_i, h_i$ corresponds to the ground truth bounding box center coordinates, width and height
for objects in grid cell (if it exists) and $\hat{x_i}, \hat{y_i}, \hat{w_i}, \hat{h_i}$ stand for the corresponding predictions. $C_i$ and $\hat{C_i}$ denote confidence score of \emph{objectness} at grid cell $i$ for ground truth and prediction. $p_i(c
)$ and $\hat{p_i}(c)$ stand for conditional probability for object class $c$ at cell index $i$ for ground truth and prediction respectively. We use similar settings for YOLO's object detection loss minimization and use values of $\lambda_{coord} = 5$
and $\lambda_{noobj} = 0.5$. \vspace{-2.5mm} \subsubsection{Similarity between \emph{Pseudo-labels} and Predictions} Our objective function also includes a regularizer that penalizes the dissimilarity between \emph{pseudo-labels} and the prediction at
each time frame $t$. \vspace{-2.5mm} \begin{equation} \label{auto_enc_loss} \mbox{similarity\_loss} = \sum^T_{t=0}\sum^{S^2}_{i=0}\hat{C}^{(t)}_i\Big(\mathbf{x}^{(t)}_i - \hat{\mathbf{y}_i}^{(t)} \Big)^2 \end{equation} Here, $\mathbf{x}^{(t)}_i$ and $\
hat{\mathbf{y}_i}^{(t)}$ denote the \emph{pseudo-labels} and predictions corresponding to the $i$-th grid cell at $t$-th time step respectively. We perform minimization of the square loss weighted by the predicted confidence score at the corresponding cell
. \subsubsection{Object Category-level Weak-Supervision} Replication of the static target at each sequential step has been shown to be effective in \cite{LiptonKEW15, yue2015beyond, dai2015semi}. Of course, with video data, different objects may move i
n different directions and speeds. Yet, within a short time duration, we could expect all objects to be present. Thus we employ target replication for classification but not localization objectives. We minimize the square loss between the categories a
ggregated over all grid cells in the ground truth $\mathbf{y}^{(T)}$ at final time step $T$ and predictions $\hat{\mathbf{y}}^{(t)}$ at all time steps $t$. Aggregated category from the ground truth considers only the cell indices where an object is p
resent. For predictions, contribution of cell $i$ is weighted by its predicted confidence score $\hat{C}^{(t)}_i$. Note that cell indices with positive detection are sparse. Thus, we consider the confidence score of each cell while minimizing the aggr
egated category loss. \vspace{-2.5mm} \begin{equation} \label{category_supervision} \mbox{category\_loss} = \sum^T_{t=0}\bigg(\sum_{c \in classes} \Big(\sum^{S^2}_{i=0} \hat{C}^{(t)}_i\big(\hat{p}^{(t)}_i(c)\big) - \sum^{S^2}_{i=0}\mathbbm{1}^{obj^{(T)}}
_i \big(p_i^{(T)}(c)\big)\Big) \bigg)^2 \end{equation} \subsubsection{Consecutive Prediction Smoothness} Additionally, we regularize the model by encouraging smoothness of predictions across consecutive time-steps. This makes sense intuitively because
we assume that objects rarely move rapidly from one frame to another. \vspace{-2.5mm} \begin{equation} \label{prediction_smoothness} \mbox{prediction\_consistency\_loss} = \sum^{T-1}_{t=0}\Big(\hat{\mathbf{y}_i}^{(t)} - \hat{\mathbf{y}_i}^{(t+1)} \Big
)^2 \end{equation} \vspace{-2.5mm} \subsection{Inference} The recurrent neural network predicts output at every time-step. The network predicts $98$ bounding boxes per video frame and class probabilities for each of the $49$ grid cells. We note that
for every cell, the net predicts class conditional probabilities for each one of the $C$ categories and $B$ bounding boxes. Each one of the $B$ predicted bounding boxes per cell has an associated \emph{objectness} confidence score. The predicted co
nfidence score at that grid is the maximum among the boxes. The bounding box with the highest score becomes the \emph{responsible} prediction for that grid cell $i$. The product of class conditional probability $\hat{p}^{(t)}_i(c)$ for category ty
pe $c$ and \emph{objectness} confidence score $\hat{C}^{(t)}_i$ at grid cell $i$, if above a threshold, infers a detection. In order for an object of category type $c$ to be detected for $i$-th cell at time-step $t$, both the class conditional probabi
lity $\hat{p}^{(t)}_i(c)$ and \emph{objectness score} $\hat{C}^{(t)}_i$ must be reasonably high. Additionally, we employ Non-Maximum Suppression (NMS) to winnow multiple high scoring bounding boxes around an object instance and produce a single detectio
n for an instance. By virtue of YOLO-style prediction, NMS is not critical. \section{Experimental Results} \label{sec:results} In this section, we empirically evaluate our model on the popular \textbf{Youtube-Objects} dataset, providing both quantit
ative results (as measured by mean Average Precision) and subjective evaluations of the model's performance, considering both successful predictions and failure cases. The \textbf{Youtube-Objects} dataset\cite{youtube-Objects} is composed of videos co
llected from Youtube by querying for the names of 10 object classes of the PASCAL VOC Challenge. It contains 155 videos in total and between 9 and 24 videos for each class. The duration of each video varies between 30 seconds and 3 minutes. However,
only $6087$ frames are annotated with $6975$ bounding-box instances. The training and test split is provided. \subsection{Experimental Setup} We implement the domain-adaption of YOLO and the proposed RNN model using Theano \cite{Theano2016arXiv160502
688short}. Our best performing RNN model uses two GRU layers of $150$ hidden units each and dropout of probability $0.5$ between layers, significantly outperforming domain-adapted YOLO alone. While we can only objectively evaluate prediction quality on
the labeled frames, we present subjective evaluations on sequences. \subsection{Objective Evaluation} We compare our approach with other methods evaluated on the Youtube-Objects dataset. As shown in Table \ref{table:per_category_results} and Table \ref{
table:final_mAP}, Deformable Parts Model (DPM) \cite{FelzenszwalbMR_CVPR_2008})-based detector reports \cite{KalogeitonFS15} mean average precision below $30$, with especially poor performance in some categories such as \emph{cat}. The method of Tripat
hi \emph{et al}\bmvaOneDot (VPO) \cite{Tripathi_WACV16} uses consistent video object proposals followed by a domain-adapted AlexNet classifier (5 convolutional layer, 3 fully connected) \cite{AlexNet12} in an R-CNN \cite{RCNN_girshick14CVPR}-like framewo
rk, achieving mAP of $37.41$. We also compare against YOLO ($24$ convolutional layers, 2 fully connected layers), which unifies the classification and localization tasks, and achieves mean Average Precision over $55$. In our method, we adapt YOLO to gen
erate \emph{pseudo-labels} for all video frames, feeding them as inputs to the refinement RNN. We choose YOLO as the \emph{pseudo-labeler} because it is the most accurate among feasibly fast image-level detectors. The domain-adaptation improves YOLO's pe
rformance, achieving mAP of $61.66$. Our model with RNN-based prediction refinement, achieves superior aggregate mAP to all baselines. The RNN refinement model using both input-output similarity, category-level weak-supervision, and prediction smoothn
ess performs best, achieving $\mbox{68.73}$ mAP. This amounts to a relative improvement of $\mbox{11.5\%}$ over the best baselines. Additionally, the RNN improves detection accuracy on most individual categories (Table \ref{table:per_category_results}).
\begin{table} \label{table:per_category_results} \centering \footnotesize \begin{tabular}{lllllllllll} \multicolumn{11}{c}{\textbf{Average Precision on 10-categories}} \\ \midrule Methods & airplane & bird & boat & car & cat & cow & dog & horse & mbi
ke & train \\ \midrule DPM\cite{FelzenszwalbMR_CVPR_2008} & 28.42 & 48.14 & 25.50 & 48.99 & 1.69 & 19.24 & 15.84 & 35.10 & 31.61 & 39.58 \\ VOP\cite{Tripathi_WACV16} & 29.77 & 28.82 & 35.34 & 41.00 & 33.7 & 57.56 & 34.42 & 54.52 & 29.77 & 29.23 \\
YOLO\cite{YOLO_RedmonDGF15} & 76.67 & 89.51 & 57.66 & 65.52 & 43.03 & 53.48 & 55.81 & 36.96 & 24.62 & 62.03 \\ DA YOLO & \textbf{83.89} & \textbf{91.98} & 59.91 & 81.95 & 46.67 & 56.78 & 53.49 & 42.53 & 32.31 & 67.09 \\ \midrule RNN-IOS & 82.7
8 & 89.51 & 68.02 & \textbf{82.67} & 47.88 & 70.33 & 52.33 & 61.52 & 27.69 & \textbf{67.72} \\ RNN-WS & 77.78 & 89.51 & \textbf{69.40} & 78.16 & 51.52 & \textbf{78.39} & 47.09 & 81.52 & 36.92 & 62.03 \\ RNN-PS & 76.11 & 87.65 & 62.16 & 80.69 & \tex
tbf{62.42} & 78.02 & \textbf{58.72} & \textbf{81.77} & \textbf{41.54} & 58.23 \\ \bottomrule \end{tabular} \caption{Per-category object detection results for the Deformable Parts Model (DPM), Video Object Proposal based AlexNet (VOP), image-trained YOLO
(YOLO), domain-adapted YOLO (DA-YOLO). RNN-IOS regularizes on input-output similarity, to which RNN-WS adds category-level weak-supervision, to which RNN-PS adds a regularizer encouraging prediction smoothness.} \end{table} \begin{table}[h] \label{t
able:final_mAP} \centering \begin{tabular}{llllllll} \multicolumn{8}{c}{\textbf{mean Average Precision on all categories}} \\ \midrule Methods & DPM & VOP & YOLO & DA YOLO & RNN-IOS & RNN-WS & RNN-PS\\ \midrule mAP & 29.41 & 37.41 & 56.53 & \tex
tbf{61.66} & 65.04 & 67.23 & \textcolor{blue}{\textbf{68.73}}\\ \bottomrule \end{tabular} \caption{Overall detection results on Youtube-Objects dataset. Our best model (RNN-PS) provides $7\%$ improvements over DA-YOLO baseline.} \end{table} \vspace{-2.5m
m} \vspace{-2.5mm} \subsection{Subjective Evaluation} We provide a subjective evaluation of the proposed RNN model in Figure \ref{fig:subjective1}. Top and bottom rows in every pair of sequences correspond to \emph{pseudo-labels} and results from our ap
proach respectively. While only the last frame in each sequence has associated ground truth, we can observe that the RNN produces more accurate and more consistent predictions across time frames. The predictions are consistent with respect to classific
ation, localization and confidence scores. In the first example, the RNN consistently detects the \emph{dog} throughout the sequence, even though the \emph{pseudo-labels} for the first two frames were wrong (\emph{bird}). In the second example, \emph
{pseudo-labels} were \emph{motorbike}, \emph{person}, \emph{bicycle} and even \emph{none} at different time-steps. However, our approach consistently predicted \emph{motorbike}. The third example shows that the RNN consistently predicts both of the ca
rs while the \emph{pseudo-labeler} detects only the smaller car in two frames within the sequence. The last two examples show how the RNN increases its confidence scores, bringing out the positive detection for \emph{cat} and \emph{car} respectively b
oth of which fell below the detection threshold of the \emph{pseudo-labeler}. \begin{figure*} \begin{center} \includegraphics[scale=0.75]{result2_category_consistency-eps-converted-to.pdf} \includegraphics[scale=0.75]{result3_category_consistency-eps
-converted-to.pdf} \includegraphics[scale=0.75]{result1_localizations-eps-converted-to.pdf} \includegraphics[scale=0.75]{result8_detection_through_consistency-eps-converted-to.pdf} \includegraphics[scale=0.75]{result16_detection_through_consist
ency-eps-converted-to.pdf} \end{center} \caption{ Object detection results from the final eight frames of five different test-set sequences. In each pair of rows, the top row shows the \emph{pseudo-labeler} and the bottom row
shows the RNN. In the first two examples, the RNN consistently predicts correct categories \emph{dog} and \emph{motorbike}, in contrast to the inconsistent baseline. In the third sequence, the RNN correctly predicts multiple instances
while the \emph{pseudo-labeler} misses one. For the last two sequences, the RNN increases the confidence score, detecting objects missed by the baseline. } \label{fig:subjective1} \end{figure*} \subsection{Areas For Improvement} The YOLO
scheme for unifying classification and localization \cite{YOLO_RedmonDGF15} imposes strong spatial constraints on bounding box predictions since each grid cell can have only one class. This restricts the set of possible predictions, which may be undes
irable in the case where many objects are in close proximity. Additionally, the rigidity of the YOLO model may present problems for the refinement RNN, which encourages smoothness of predictions across the sequence of frames. Consider, for example, an
object which moves slightly but transits from one grid cell to another. Here smoothness of predictions seems undesirable. \begin{figure*} \begin{center} \includegraphics[scale=0.75]{failure_cases-eps-converted-to.pdf} \end{center} \caption{Failure ca
ses for the proposed model. Left: the RNN cannot recover from incorrect \emph{pseudo-labels}. Right: RNN localization performs worse than \emph{pseudo-labels} possibly owing to multiple instances of the same object. } \label{fig:failure_cases} \vspace{-2.
5mm} \end{figure*} Figure \ref{fig:failure_cases} shows some failure cases. In the first case, the \emph{pseudo-labeler} classifies the instances as \emph{dogs} and even as \emph{birds} in two frames whereas the ground truth instances are \emph
{horses}. The RNN cannot recover from the incorrect pseudo-labels. Strangely, the model increases the confidence score marginally for a different wrong category \emph{cow}. In the second case, possibly owing to motion and close proximity of multiple inst
ances of the same object category, the RNN predicts the correct category but fails on localization. These point to future work to make the framework robust to motion. The category-level weak supervision in the current scheme assumes the presence of al
l objects in nearby frames. While for short snippets of video this assumption generally holds, it may be violated in case of occlusions, or sudden arrival or departure of objects. In addition, our assumptions regarding the desirability of prediction sm
oothness can be violated in the case of rapidly moving objects. \vspace{-2.5mm} \section{Related Work} \label{sec:prior-art} Our work builds upon a rich literature in both image-level object detection,video analysis, and recurrent neural networks. Seve
ral papers propose ways of using deep convolutional networks for detecting objects \cite{RCNN_girshick14CVPR,fast_RCNN_15,Faster_RCNN_RenHG015, YOLO_RedmonDGF15, SzegedyREA14, Inside_Outside_Net_BellZBG15, DeepID-Net_2015_CVPR, Overfeat_SermanetEZMFL13, CR
AFTCVPR16, Gidaris_2015_ICCV}. Some approaches classify the proposal regions \cite{RCNN_girshick14CVPR,fast_RCNN_15} into object categories and some other recent methods \cite{Faster_RCNN_RenHG015, YOLO_RedmonDGF15} unify the localization and classificati
on stages. Kalogeiton \emph{et al}\bmvaOneDot \cite{KalogeitonFS15} identifies domain shift factors between still images and videos, necessitating video-specific object detectors. To deal with shift factors and sparse object-level annotations in video,
researchers have proposed several strategies. Recently, \cite{Tripathi_WACV16} proposed both transfer learning from the image domain to video frames and optimizing for temporally consistent object proposals. Their approach is capable of detecting both
moving and static objects. However, the object proposal generation step that precedes classification is slow. Prest \emph{et al}\bmvaOneDot \cite{Weak_obj_from_videoPrestLCSF12}, utilize weak supervision for object detection in videos via category-le
vel annotations of frames, absent localization ground truth. This method assumes that the target object is moving, outputting a spatio-temporal tube that captures this most salient moving object. This paper, however, does not consider context within video
for detecting multiple objects. A few recent papers \cite{DeepID-Net_2015_CVPR, Inside_Outside_Net_BellZBG15} identify the important role of context in visual recognition. For object detection in images, Bell \emph{et al}\bmvaOneDot \cite{Inside_Outside_
Net_BellZBG15} use spatial RNNs to harness contextual information, showing large improvements on PASCAL VOC \cite{PASCAL_VOC} and Microsoft COCO \cite{COCOLinMBHPRDZ14} object detection datasets. Their approach adopts proposal generation followed by cla
ssification framework. This paper exploits spatial, but not temporal context. Recently, Kang \emph{et al}\bmvaOneDot \cite{KangCVPR16} introduced tubelets with convolutional neural networks (T-CNN) for detecting objects in video. T-CNN uses spatio-tem
poral tubelet proposal generation followed by the classification and re-scoring, incorporating temporal and contextual information from tubelets obtained in videos. T-CNN won the recently introduced ImageNet object-detection-from-video (VID) task with
provided densely annotated video clips. Although the method is effective for densely annotated training data, it's behavior for sparsely labeled data is not evaluated. By modeling video as a time series, especially via GRU \cite{Cho14_GRU} or LSTM RNN
s\cite{LSTM_Hochreiter_97}, several papers demonstrate improvement on visual tasks including video classification \cite{yue2015beyond}, activity recognition \cite{LongTermRecurrentDonahueHGRVSD14}, and human dynamics \cite{Fragkiadaki_2015_ICCV}. These
models generally aggregate CNN features over tens of seconds, which forms the input to an RNN. They perform well for global description tasks such as classification \cite{yue2015beyond,LongTermRecurrentDonahueHGRVSD14} but require large annotated dataset
s. Yet, detecting multiple generic objects by explicitly modeling video as an ordered sequence remains less explored. Our work differs from the prior art in a few distinct ways. First, this work is the first, to our knowledge, to demonstrate the capa
city of RNNs to improve localized object detection in videos. The approach may also be the first to refine the object predictions of frame-level models. Notably, our model produces significant improvements even on a small dataset with sparse annotations.
\vspace{-2.5mm} \vspace{-2.5mm} \section{Conclusion} We introduce a framework for refining object detection in video. Our approach extracts contextual information from neighboring frames, generating predictions with state of the art accuracy
that are also temporally consistent. Importantly, our model benefits from context frames even when they lack ground truth annotations. For the recurrent model, we demonstrate an efficient and effective training strategy that simultaneously employs l
ocalization-level strong supervision, category-level weak-supervision, and a penalty encouraging smoothness of predictions across adjacent frames. On a video dataset with sparse object-level annotation, our framework proves effective, as validated by
extensive experiments. A subjective analysis of failure cases suggests that the current approach may struggle most on cases when multiple rapidly moving objects are in close proximity. Likely, the sequential smoothness penalty is not optimal for suc
h complex dynamics. Our results point to several promising directions for future work. First, recent state of the art results for video classification show that longer sequences help in global inference. However, the use of longer sequences for localizat
ion remains unexplored. We also plan to explore methods to better model local motion information with the goal of improving localization of multiple objects in close pro
ximity. Another promising direction, we would like to experiment with loss functions to incorporate specialized handling of classification and localization objectives.
\section{Introduction} The wave-particle duality is an alternative statement of the complementarity principle, and it establishes the relation between \ankb{corpuscular and undulatory}{the corpuscular and the ondulatory} nature of quantum entities \cite{B
ohr1928}. It can be illustrated in a two-way interferometer, where the apparatus can be set to observe \ankb{particle}{the particle} behavior when a single path is \ankb{taken}{taken,} or wave-like behavior, when the impossibility to define a path is shown
by the interference. A modern approach to the wave-particle duality includes quantitative relations between quantities that represent the possible \textit{a priori} knowledge of the which-way information (\ankb{predictability}{predicability}) and the ``qu
ality'' of the interference fringes (\ankb{Visibility}{visibility}). Several publications in the literature \cite{Bohr1928, Wootters1979, Summhammer1987, Greenberger1988, Mandel1991} contributed to the formulation of the quantitative analysis of the wave-p
article duality. For a bipartite system\ankb{ entanglement, the quantum correlations between each part, can play a role. Such correlations can}{, entanglement can} give an extra which-way (path) information about the interferometric possibilities. The quan
titative relations for systems composed by two particles were \ankb{}{extensively} studied in \cite{Jaeger1993, Jaeger1995, Englert1996, Englert2000, Scully1989, Scully1991, Mandel1995, Tessier2005, Jakob2010, Miatto2015,Bagan2016, Coles2016}. Therefore, \
ankb{understand}{understanding} the behavior of such quantities, in various regimes and situations, is essential to answer fundamental and/or technological questions of the quantum theory \cite{Greenberger1999}. \alams{The Complementarity quantities can p
resent interesting dynamical behaviors,}{Concerning the study of the dynamical behavior of complementarity quantities,} an example is the so-called \textit{quantum eraser}, \alams{where an increase or preservation of the visibility of an interferometer exp
eriment is caused when the ``which-way'' information is erased.}{\ankb{ i.e. an increasing or preservation of the \ankb{Visibility}{visibility} in an interferometric scheme (or the ``erasure'' of the which-way information probably stored in the initial sta
te).}{ where an increase or preservation of the visibility of an interferometer experiment is caused when the which-path information is erased.}} Since its proposal \cite{Scully1982}\ankb{ this phenomena}{, it} has \ankb{investigated it carefully,}{been in
vestigated carefully} both theoretically and experimentally (see for example Refs. \cite{Englert2000, Scully1991, Mandel1995, Storey1994, Wiseman1995, Mir2007, Luis1998, Busch2006, Rossi2013, Walborn2002, Mir2007, Teklemariam2001, Teklemariam2002, Kim2000,
Salles2008, Heuer2015}). In \ankb{a recent work }{Ref.~}\cite{Rossi2013}, the authors explore the quantum eraser problem in \ankb{multipartite}{a multipartite} model\ankb{, where two cavities ($q_A+q_B$), which will be taken as a two qubit system $A + B$,
in an initial maximally entangled state (and therefore with zero \ankb{Visibility}{visibility}), couple through a Jaynes-Cummings Hamiltonian to $N$ two-level atoms (we will call the global system as $q_A + q_B + R$, where all the individual systems are q
ubits)}{. Initially a bipartite qubit system is prepared in a maximally entangled state and interacts with $N$ other qubits. This model can be implemented considering the qubits of interest the cavity modes of two cavities and the $N$ qubits as two-level a
toms}. In this work \cite{Rossi2013}, an increase of visibility is achieved by performing appropriate projective measurements. An intrinsic relation between the complementarity quantities and the performed measurements is outlined: since \ankb{they}{the me
asurements} were made in order to obtain an \ankb{increasing}{increase} of the \ankb{Visibility}{visibility}, the remaining quantities (Entanglement as measured by the concurrence, and the predictability) must obey a ``complementary'' behavior. In that cas
e, \ankb{Visibility}{visibility} and predictability increases, and entanglement decreases, since the measurements are made in order to \ankb{establishes}{establish} the quantum eraser. In Reference \cite{Rossi2013} only the maximization of the visibility w
as considered, in the present work we extend the analysis and consider maximization of predictability, visibility and concurrence. Also, in the previous work \cite{Rossi2013}, only one value of the coupling constant was considered. In this contribution we