Online Tracking by Learning Discriminative Saliency Map with Convolutional Neural Network

Introduction

Object tracking has played important roles in a wide range of computer vision applications. Although it has been studied extensively during past decades, object tracking is still a difficult problem due to many challenges in real world videos such as occlusion, pose variations, illumination changes, fast motion, and background clutter. Success in object tracking relies heavily on how robust the representation of target appearance is against such challenges.

For this reason, reliable target appearance modeling problem has been investigated in recent tracking algorithms actively (Bao et al., 2012; Jia et al., 2012; Mei & Ling, 2009; Zhang et al., 2012; Zhong et al., 2012; Ross et al., 2004; Han et al., 2008; Babenko et al., 2011; Hare et al., 2011; Grabner et al., 2006; Saffari et al., 2010), which are classified into two major categories depending on learning strategies: generative and discriminative methods. In generative framework, the target appearance is typically described by a statistical model estimated from tracking results in previous frames. To maintain the target appearance model, various approaches have been proposed including sparse representation (Bao et al., 2012; Jia et al., 2012; Mei & Ling, 2009; Zhang et al., 2012; Zhong et al., 2012), online density estimation (Han et al., 2008), incremental subspace learning (Ross et al., 2004), etc. On the other hand, discriminative framework (Babenko et al., 2011; Hare et al., 2011; Grabner et al., 2006; Saffari et al., 2010) aims to learn a classifier that discriminates target from surrounding background. Various learning algorithms have been incorporated including online boosting (Grabner et al., 2006; Saffari et al., 2010), multiple instance learning (Babenko et al., 2011), structured support vector machine (Hare et al., 2011), and online random forest (Gall et al., 2011; Schulter et al., 2011). These approaches are limited to using too simple and/or hand-crafted features for target representation, such as template, Haar-like features, histogram features and so on, which may not be effective to handle latent challenges imposed on video sequences.

Convolutional Neural Network (CNN) has recently drawn a lot of attention in computer vision community due to its representation power. (Krizhevsky et al., 2012) trained a network using 1.2 million images for image classification and demonstrated significantly improved performance in ImageNet challenge (Berg et al., 2012). Since the huge success of this work, CNN has been applied to representing images or objects in various computer vision tasks including object detection (Girshick et al., 2014; Sermanet et al., 2014; He et al., 2014), object recognition (Oquab et al., 2014; Donahue et al., 2014; Zhang et al., 2014), pose estimation (Toshev & Szegedy, 2014), image segmentation (Hariharan et al., 2014), image stylization (Karayev et al., 2014), etc.

Despite such popularity, there are only few attempts to employ CNNs for visual tracking since offline classifiers are not appropriate for visual tracking conceptually and online learning based on CNN is not straightforward due to large network size and lack of training data. In addition, the feature extraction from the deep structure may not be appropriate for visual tracking because the visual features extracted from top layers encode semantic information and exhibit relatively poor localization performance in general. (Fan et al., 2010) presents a human tracking algorithm based on a network trained offline, but it needs to learn a separate class-specific network to track other kind of objects. On the other hand, (Li et al., 2014) proposes a target-specific CNN for object tracking, where the CNN is trained incrementally during tracking with new examples obtained online. The network used in this work is shallow since learning a deep network using a limited number of training examples is challenging, and the algorithm fails to take advantage of rich information extracted from deep CNNs. There is a tracking algorithm based on a pre-trained network (Wang & Yeung, 2013), where a stacked denoising autoencoder is trained using a large number of images to learn generic image features. Since this network is trained with tiny gray images and has no shared weight, its representation power is limited compared to recently proposed CNNs.

We propose a novel tracking algorithm based on a pre-trained CNN to represent target, where the network is trained originally for large-scale image classification. On top of the hidden layers in the CNN, we put an additional layer of an online Support Vector Machine (SVM) to learn a target appearance discriminatively against background. The model learned by SVM is used to compute a target-specific saliency map by back-propagating the information relevant to target to input layer (Simonyan et al., 2014). We exploit the target-specific saliency map to obtain generative target appearance models (filters) and perform tracking with understanding of spatial configuration of target. The overview of our algorithm is illustrated in Figure 1, and the contributions of this paper are summarized below:

Although recent tracking methods based on CNN typically attempt to learn a network in an online manner (Li et al., 2014), our algorithm employs a pre-trained CNN to represent generic objects for tracking and achieves outstanding performance empirically.

We propose a technique to construct a target-specific saliency map by back-propagating only relevant features through CNN, which overcomes the limitation of the existing method to visualize saliency corresponding to the predefined classes only. This technique also enable us to obtain pixel-level target segmentation.

We learn a simple target-specific appearance filter online and apply it to the saliency map; this strategy improves target localization performance even with shift-invariant property of CNN-based features.

The rest of this paper is organized as follows. We first describe the overall framework of our algorithm in Section 2 and the detailed methodology is discussed in Section 3. The performance of our algorithm is presented in Section 4.

Overview of Our Algorithm

Our tracking algorithm employs a pre-trained CNN to represent target. In each frame, it first draws samples for candidate bounding boxes near the target location in the previous frame, takes their image observations, and extracts feature descriptors for the samples using the pre-trained CNN. We found out that the features from the CNN capture semantic information of target effectively and handle various geometric and photometric transformations successfully as reported in (Oquab et al., 2014; Karayev et al., 2014; Donahue et al., 2014). However, it may lose some spatial information of the target due to pooling operations in CNN, which is not desirable for tracking since the spatial configuration is a useful cue for accurate target localization.

To fully exploit the representation power of CNN features while preserving spatial information of target, we adopt the target-specific saliency map as our observation for tracking, which is generated by back-propagating target-specific information of CNN features to input layer. This technique is inspired by (Simonyan et al., 2014), where class-specific saliency map is constructed by back-propagating the information corresponding to the identified label to visualize the region of interest. Since target in visual tracking problem belongs to an arbitrary class and its label is unknown in advance, the model for target class is hard to be pre-trained.

Hence, we employ an online SVM, which discriminates target from background by learning target-specific information in the CNN features; the target-specific information learned by the online SVM can be regarded as label information in the context of (Simonyan et al., 2014). The SVM classifies each sample, and we compute the saliency map for each positive example by back-propagating its CNN feature along the pre-trained CNN with guidance of the SVM till the input layer. Each saliency map highlights regions discriminating target from background. The saliency maps of the positive examples are aggregated to build the target-specific saliency map. The target-specific saliency map alleviates the limitation of CNN features for tracking by providing important spatial configuration of target.

Our tracking algorithm is then formulated as a sequential Bayesian filtering framework using the target-specific saliency map for observation in tracking. A generative appearance model is constructed by accumulating target observations in target-specific saliency maps over time, which reveals meaningful spatial configuration of target such as shape and parts. A dense likelihood map of each frame is computed efficiently by convolution between the target-specific saliency map and the generative appearance model. The overall algorithm is illustrated in Figure 1.

Our algorithm exploits the discriminative properties of online SVM, which helps generate target-specific saliency map. In addition, we construct the generative appearance model from the saliency map and perform tracking through sequential Bayesian filtering. This is a natural combination of discriminative and generative approaches, and we take the benefits from both frameworks.

Proposed Algorithm

This section describes the comprehensive procedure of our tracking algorithm. We first discuss the features obtained from pre-trained CNN. The method to construct target-specific saliency map are presented in detail, and how the saliency map can be employed for constructing generative models and tracking object is described. After that, we present online SVM technique employed to learn target appearance in a discriminative manner sequentially.

To represent target appearances, our tracking algorithm employs a CNN, which is pre-trained on a large number of images. The pre-trained generic model is useful especially for online tracking since it is not straightforward to collect a sufficient number of training data. In this paper, R-CNN (Girshick et al., 2014) is adopted as the pre-trained model, but other CNN models can be used alternatively. Out of the entire network structure, we take outputs from the first fully-connected layer as they tend to capture general characteristics of objects and have shown excellent generalization performance in many other domains as described in (Donahue et al., 2014).

For a target proposal $\mathbf{x}_{i}$, the CNN takes its corresponding image observation $\mathbf{z}_{i}$ as its input, and returns an output from the first fully-connected layer $\phi(\mathbf{x}_{i})$ as a feature vector of $\mathbf{x}_{i}$. We apply the SVM to each CNN feature vector $\phi(\mathbf{x}_{i})$ and classify $\mathbf{x}_{i}$ into either positive or negative.

2 Target-Specific Saliency Map Estimation

For target tracking, we first compute SVM scores of candidate samples represented by the CNN features and classify them into target or background. Based on this information, one naïve option to complete tracking is to simply select the optimal sample with the maximum score as

However, this approach typically has the limitation of inaccurate target localization since, when calculating $\phi(\mathbf{x}_{i})$, the spatial configuration of target may be lost by spatial pooling operations (Fan et al., 2010).

To handle the localization issue while enjoying the effectiveness of CNN features, we propose the target-specific saliency map, which highlights discriminative target regions within the image. This is motivated by the class-specific saliency map discussed in (Simonyan et al., 2014). The class-specific saliency map of a given image $I$ is the gradient of class score $S_{c}(I)$ with respect to the image as

The saliency map is constructed by back-propagation. Specifically, let $f^{(1)},\dots,f^{(L)}$ and $F^{(1)},\dots,F^{(L)}$ denote the transformation functions and their outputs in the network, where $F^{(l)}=f^{(l)}\circ f^{(l-1)}\circ\cdots\circ f^{(1)}(x)$ and $S_{c}(I)=F^{(L)}$. Eq. (1) is computed using chain rule as

Intuitively, the pixels that are closely related to the class $c$ affect changes in $S_{c}$ more, which means that nearby regions of such pixels would have high values in saliency map.

When calculating such saliency map for object tracking, we impose target-specific information instead of class membership due to the reasons discussed in Section 2. For the purpose, we adopt the SVM weight vector $\mathbf{w}=(w_{1},\dots,w_{n})^{\top}$, which is learned online to discriminate between target and background. Since the last fully-connected layer corresponds to the online SVM, the outputs of the last two layers in our network are given by

Plugging Eq. (3) and (4) into Eq. (2), the gradient map of the target proposal $\mathbf{x}_{i}$ is given by

where $\mathbf{z}_{i}$ is the image observation of $\mathbf{x}_{i}$.

Instead of using all entries in $\phi(\mathbf{x}_{i})$ to generate target-specific saliency map, we only select the dimensions corresponding to positive weights in $\mathbf{w}$ since they have clearer contribution to make $\mathbf{x}_{i}$ positive. Note that every element in $\phi(\mathbf{x}_{i})$ is positive due to ReLU operations in CNN learning. Then, we obtain the target-specific feature $\phi^{+}(\mathbf{x}_{i})$ as

where $\phi_{k}(\mathbf{x}_{i})$ denotes the $k$-th entry of $\phi(\mathbf{x}_{i})$. Then the gradient of target-specific feature $\phi^{+}(\mathbf{x}_{i})$ with respect to the image observation is obtained by

Since the gradient is computed only for the target-specific information $\phi^{+}(\mathbf{x}_{i})$, pixels to distinguish the target from background would have high values in $g(\mathbf{x}_{i})$.

The target-specific saliency map $M$ is obtained by aggregating $g(\mathbf{x}_{i})$ of samples with positive SVM scores in image space. As $g(\mathbf{x}_{i})$ is defined over sample observation $\mathbf{z}_{i}$, we first project it to image space and zero-pad outside of $\mathbf{z}_{i}$; we denote the result by $G_{i}$ afterwards. Then, the target-specific saliency map is obtained by taking the pixelwise maximum magnitude of the gradient maps $G_{i}$’s corresponding to positive examples, which is given by

where $p$ denotes pixel location. We suppress erroneous activations from background by considering only positive examples when aggregating sample gradient maps. An example of target-specific saliency map is illustrated in Figure 2, where strong activations typically come from target areas and spatial layouts of target are exposed clearly.

3 Target Localization with Saliency Map

Given the target-specific saliency map at frame $t$ denoted by $M_{t}$, the next step of our algorithm is to locate the target through sequential Bayesian filtering. Let $\mathbf{x}_{t}$ and $M_{t}$ denote the state and observation variables at current frame $t$, respectively, where saliency map is used for measurement. The posterior of the target state $p(\mathbf{x}_{t}|M_{1:t})$ is given by

where $p(\mathbf{x}_{t}|M_{1:t-1})$ denotes the prior distribution predicted from the previous time step, and $p(M_{t}|\mathbf{x}_{t})$ means observation likelihood.

The prior distribution $p(\mathbf{x}_{t}|M_{1:t-1})$ of target state at the current time step is estimated from the posterior at the previous frame through prediction, which is given by

where $p(\mathbf{x}_{t}|\mathbf{x}_{t-1})$ denotes a state transition model. Target dynamics between two consecutive frames is given by a simple linear equation as

where $\mathbf{d}_{t}$ denotes a displacement of target location, and $\boldsymbol{\varepsilon}_{t}$ indicates a Gaussian noise. Both $\mathbf{d}_{t}$ and $\boldsymbol{\varepsilon}_{t}$ are unknown before tracking in general, but is estimated from the samples classified as target by our online SVM in our case. Specifically, $\mathbf{d}_{t}$ and $\boldsymbol{\varepsilon}_{t}$ are given respectively by

where $\mathbf{x}_{t-1}^{*}$ denotes the target location at the previous frame, and $\boldsymbol{\mu}_{t}$ and $\Sigma_{t}$ indicate mean and variance of locations of positive samples at the current frame, respectively. From Eq. (13) and (14), the transition model for prediction is derived as follows:

Since the transition model is linear with Gaussian noise, computation of the prior in Eq. (12) can be performed efficiently by transforming the posterior $p(\mathbf{x}_{t-1}|M_{1:t-1})$ at the previous step by $\mathbf{d}_{t}$ and applying Gaussian smoothing with covariance $\Sigma_{t}$.

The measurement density function $p(M_{t}|\mathbf{x}_{t})$ represents the likelihood in the state space, which is typically obtained by computing the similarity between the appearance models of target and candidates. In our case, we utilize $M_{t}$, target-specific saliency map at frame $t$, for observation to compute the likelihood of each target state. Note that pixel-wise intensity and its spatial configuration in the saliency map provide useful information for target localization. At frame $t$, we construct the target appearance model $H_{t}$ given the previous saliency maps $M_{1:t-1}$ in a generative way. Let $M_{k}(\mathbf{x}_{k}^{*})$ denote the target filter at frame $k$, which is obtained by extracting the subregion in $M_{k}$ at the location corresponding to the optimal target bounding box given by $\mathbf{x}^{*}_{k}$. The appearance model $H_{t}$ is constructed by aggregating the recent target filters as follows:

where $m$ is a constant for the number of target filters to be used for model construction. The main idea behind Eq. (16) is that the local saliency map nearby the optimal target location in a frame plays a role as a filter to identify the target within the saliency map in the subsequent frames. Since the target filter is computed based on $m$ recent filters, we need to store the $m$ filters to update the target filter. Therefore, given the appearance model defined in Eq. (16), the observation likelihood $p(M_{t}|\mathbf{x}_{t})$ is computed by simple convolution between $H_{t}$ and $M_{t}$ by

where $\otimes$ denotes convolution operator. This is similar to the procedure in object detection, e.g., (Felzenszwalb et al., 2010), where the filter is constructed from features to represent the object category and applied to the feature map to localize the object by convolution.

Given the prior in Eq. (12) and the likelihood in Eq. (17), the target posterior at the current frame is computed simply by applying Eq. (11). Once the target posterior is obtained, the optimal target state is given by solving the maximum a posteriori problem as

Once tracking at frame $t$ is completed, we update the classifier based on $\mathbf{x}_{t}^{*}$, which is discussed next.

4 Discriminative Model Update by Online SVM

We employ an online SVM to learn a discriminative model of target. Our SVM can be regarded as a fully-connected layer with a single node but provides a fast and exact solution in a single pass to learn a model incrementally.

Given a set of samples with associated labels, $\{(\mathbf{x}^{\prime}_{i},y^{\prime}_{i})\}$, obtained from the current tracking results, we hope to update a weight vector $\mathbf{w}$ of SVM. The label $y^{\prime}_{i}$ of a new example $\mathbf{x}^{\prime}_{i}$ is given by

where $\mathsf{BB}(\mathbf{x})$ denotes the bounding box corresponding to the given state $\mathbf{x}$ and $\delta$ denotes a pre-defined threshold. Note that the examples with the bounding box overlap ratios larger than $\delta$ are not included in the training set for our online learning to avoid drift problem.

Before discussing online SVM, we briefly review the optimization procedure of an offline learning algorithm. Given training examples $\{(\mathbf{x}_{i},y_{i})\}$, the offline SVM learns a weight vector $\mathbf{w}=(w_{1},\dots,w_{n})^{\top}$ by solving a quadratic convex optimization problem. The dual form of SVM objective function is given by

where $\{a_{i}\}$ are Largrange multipliers, $b$ is bias, and $Q_{ij}=y_{i}y_{j}K(\mathbf{x}_{i},\mathbf{x}_{j})$. In our tracking algorithm, the kernel function is defined by the inner product between two CNN features, i.e., $K(\mathbf{x}_{i},\mathbf{x}_{j})=\phi(\mathbf{x}_{i})^{\top}\phi(\mathbf{x}_{j})$. In online tracking, it is not straightforward for conventional QP solvers to handle the optimization problem in Eq. (22) as training data are given sequentially, not at once. Incremental SVM (Diehl & Cauwenberghs, 2003; Cauwenberghs & Poggio, 2000) is an algorithm designed to learn SVMs in such cases. The key idea of the algorithm is to retain KKT conditions on all the existing examples while updating model with a new example, so that it guarantees an exact solution at each increment of dataset. Specifically, KKT conditions are the first-order necessary conditions for the optimal solution of Eq. (22), which are given by

where $\frac{\partial W}{\partial a_{i}}$ is related to the margin of the $i$-th example that is denoted by $m_{i}$ afterwards. By the conditions in Eq. (26), each training example belongs to one of the following three categories: $E_{1}$ for support vectors lying on the margin ($m_{i}=0$), $E_{2}$ for support vectors inside the margin ($m_{i}<0$), and $E_{3}$ for non-support vectors.

Given the $k$-th example, incremental SVM estimates its Lagrangian multiplier $a_{k}$ while retaining the KKT conditions on all the existing $k-1$ training examples. In a nutshell, $a_{k}$ is initialized to 0 and updated by increasing its value over iterations. In each iteration, the algorithm estimates the largest possible increment $\Delta a_{k}$ that guarantees KKT conditions on the existing examples, and updates $a_{k}$ and existing model parameters with $\Delta a_{k}$. This iterative procedure will stop when the $k$-th example becomes a support vector or at least one existing example changes its membership across $E_{1}$, $E_{2}$, and $E_{3}$. We can generalize this online update procedure easily when multiple examples are provided as new training data. With the new and updated Lagrangian multipliers, the weight vector $\mathbf{w}$ is given by

For efficiency, we maintain only a fixed number of support vectors with smallest margins during tracking. We ask to refer to (Diehl & Cauwenberghs, 2003; Cauwenberghs & Poggio, 2000) for more details. Also, note that any other methods for online SVM learning, such as LaSVM (Bordes et al., 2005) and LaRank (Bordes et al., 2007), can also be adopted in our framework.

Experiments

This section describes our implementation details and experimental setting. The effectiveness of our tracking algorithm is then demonstrated by quantitative and qualitative analysis on a large number of benchmark sequences.

2 Analysis of Generative Appearance Models

The generative model $H_{t}$ is used to localize the target using the target-specific saliency map. As described earlier, the target-specific saliency map shows high responses around discriminative target regions; our generative model exploits such property and is constructed using the saliency maps in the previous frames. Figure 3 illustrates examples of the learned generative models in several sequences. Generally, the model successfully captures parts and shape of an object, which are useful to discriminate the target from background. More importantly, the distribution of responses within the model reveals the spatial configuration of the target, which provides a strong cue for precise localization. This can be clearly observed in examples of face and doll, where the scores from the areas of eyes and nose can be used to localize the target. When target is not rigid (e.g., person), we observe that the model has stronger responses on less deformable parts of the target (e.g., head) and localization relies more on the stable parts consequently.

3 Evaluation

To evaluate the performance, we employ all 50 sequences from the recently released tracking benchmark dataset (Wu et al., 2013). The sequences in the dataset involve various tracking challenges such as illumination variation, deformation, motion blur, background clutter, etc. We compared our method with top 10 trackers in (Wu et al., 2013), which include SCM (Zhong et al., 2012), Struck (Hare et al., 2011), TLD (Kalal et al., 2012), ASLA (Jia et al., 2012), CXT (Dinh et al., 2011), VTD (Kwon & Lee, 2010), VTS (Kwon & Lee, 2011), CSK (Henriques et al., 2012), LSK (Liu et al., 2011) and DFT (Sevilla-Lara & Learned-Miller, 2012). We used the reported results in (Wu et al., 2013) for these tracking algorithms.

Evaluation methodology

We follow the evaluation protocols in (Wu et al., 2013), where the performance of trackers are measured based on two different metrics: success rate and precision plots. In both metrics, the ratio of successfully tracked frames is measured by a set of thresholds, where bounding box overlap ratio and center location error are employed in success rate plot and precision plot, respectively. We rank the tracking algorithms based on Area Under Curve (AUC) for success rate plot and center location error at 20 pixels for precision plot.

Quantitative results in bounding box

We evaluate our method quantitatively and make a comparative study with other methods in all the 50 benchmark sequences; the results are summarized in Figure 4 for both of success rate and precision plots. In both measures, our method outperforms all other trackers with substantial margins. It is probably because the CNN features are more effective to represent high-level concept of target than hand-crafted ones although the network is trained offline for other purpose. We also compare our full algorithm with its reduced version denoted by $\text{Ours}_{\text{SVM}}$, which depends only on SVM scores as conventional tracking-by-detection algorithms do. Our full algorithm achieves non-trivial performance improvement over the reduced version, which shows that our generative model based on target-specific saliency map is useful to localize target in general.

To gain more insight about the proposed algorithm, we evaluate the performance of trackers based on individual attributes provided in the benchmark dataset. Note that the attributes describe 11 different types of tracking challenges and are annotated for each sequence. Table 5 and 5 summarize the results in two different measures. The numbers next to the attributes indicate the number of sequences involving the corresponding attribute. As illustrated in the tables, our algorithm consistently outperforms other methods in almost all challenges, and our full algorithm is generally better than its reduced version.

Quantitative results in segmentation The proposed algorithm produces pixel-wise target segmentation using target-specific discriminative saliency map. To evaluate segmentation accuracy, we select 9 video sequences from the online tracking benchmark datasetSince accurate annotation of segmentation is labor intensive and time consuming, we selected a subset of sequences (typically short ones) for evaluation. and annotate ground-truth segmentation for each sequence. The selected sequences cover various attributes in tracking challenges, and the list of sequences with associated attributes are summarized in Table 3.

The segmentation performance of the proposed algorithm is evaluated based on the overlap ratio—intersection over union—between ground-truth and identified target segmentation. As other trackers used for comparison may not be able to generate pixel-wise segmentation, we employ their bounding box outputs as segmentation masks and compute the overlap ratio with respect to the ground-truth segmentation. The results are presented by success plot as in Figure 6, where $\text{Ours}_{\text{seg}}$ denotes the proposed algorithm with target segmentation. According to Figure 6, our method outperforms all other trackers with substantial margin. Especially, we can observe a large performance improvement of the proposed target segmentation algorithm over our bonding box trackers denoted by Ours and $\text{Ours}_{\text{SVM}}$. It suggests that the proposed target-specific saliency map is sufficiently accurate to estimate the target area in a video thus can be utilized to further improve tracking.

Qualitative Results

We present the results of several sequences in Figure 5, where original frames with tracking results, target-specific saliency maps, and segmentation results are illustrated. We can observe that our algorithm also demonstrates superior performance to other algorithms qualitatively.

Conclusion

We proposed a novel visual tracking algorithm based on pre-trained CNN, where outputs from the last convolutional layer of the CNN are employed as generic feature descriptors of objects, and discriminative appearance models are learned online using an online SVM. With CNN features and learned discriminative model, we compute the target-specific saliency map by back-propagation, which highlights the discriminative target regions in spatial domain. Tracking is performed by sequential Bayesian filtering with the target-specific saliency map as observation. The proposed algorithm achieves substantial performance gain over the existing state-of-the-art trackers and shows the capability for target segmentation.

References