Visual target tracking is one of the most sought-after yet challenging research topics in computer vision. Given the ill-posed nature of the problem and its popularity in a broad range of real-world scenarios, a number of large-scale benchmark datasets have been established, on which considerable methods have been developed and demonstrated with significant progress in recent years - predominantly by recent deep learning (DL)-based methods. This survey aims to systematically investigate the current DL-based visual tracking methods, benchmark datasets, and evaluation metrics. It also extensively evaluates and analyzes the leading visual tracking methods. First, the fundamental characteristics, primary motivations, and contributions of DL-based methods are summarized from nine key aspects of: network architecture, network exploitation, network training for visual tracking, network objective, network output, exploitation of correlation filter advantages, aerial-view tracking, long-term tracking, and online tracking. Second, popular visual tracking benchmarks and their respective properties are compared, and their evaluation metrics are summarized. Third, the state-of-the-art DL-based methods are comprehensively examined on a set of well-established benchmarks of OTB2013, OTB2015, VOT2018, LaSOT, UAV123, UAVDT, and VisDrone2019. Finally, by conducting critical analyses of these state-of-the-art trackers quantitatively and qualitatively, their pros and cons under various common scenarios are investigated. It may serve as a gentle use guide for practitioners to weigh when and under what conditions to choose which method(s). It also facilitates a discussion on ongoing issues and sheds light on promising research directions.

Deep Learning for Visual Tracking: A Comprehensive Survey

Visual target tracking is one of the most sought-after yet challenging research topics in computer vision. Given the ill-posed nature of the problem and its popularity in a broad range of real-world scenarios, a number of large-scale benchmark datasets have been established, on which considerable methods have been developed and demonstrated with significant progress in recent years -- predominantly by recent deep learning (DL)-based methods. This survey aims to systematically investigate the current DL-based visual tracking methods, benchmark datasets, and evaluation metrics. It also extensively evaluates and analyzes the leading visual tracking methods. First, the fundamental characteristics, primary motivations, and contributions of DL-based methods are summarized from nine key aspects of: network architecture, network exploitation, network training for visual tracking, network objective, network output, exploitation of correlation filter advantages, aerial-view tracking, long-term tracking, and online tracking. Second, popular visual tracking benchmarks and their respective properties are compared, and their evaluation metrics are summarized. Third, the state-of-the-art DL-based methods are comprehensively examined on a set of well-established benchmarks of OTB2013, OTB2015, VOT2018, LaSOT, UAV123, UAVDT, and VisDrone2019. Finally, by conducting critical analyses of these state-of-the-art trackers quantitatively and qualitatively, their pros and cons under various common scenarios are investigated. It may serve as a gentle use guide for practitioners to weigh when and under what conditions to choose which method(s). It also facilitates a discussion on ongoing issues and sheds light on promising research directions. 

Hough transform based object detectors learn a mapping from the image domain to a Hough voting space. Within this space, object hypotheses are formed by local maxima. The votes contributing to a hypothesis are called support. In this work, we investigate the use of the support and its backprojection to the image domain for multi-view object detection. To this end, we create a shared codebook with training and matching complexities independent of the number of quantized views. We show that since backprojection encodes enough information about the viewpoint all views can be handled together. In our experiments, we demonstrate that superior accuracy and efficiency can be achieved in comparison to the popular one-vs-the-rest detectors by treating views jointly especially with few training examples and no view annotations. Furthermore, we go beyond the detection case and based on the support we introduce a part-based similarity measure between two arbitrary detections which naturally takes spatial relationships of parts into account and is insensitive to partial occlusions. We also show that backprojection can be used to efficiently measure the similarity of a detection to all training examples. Finally, we demonstrate how these metrics can be used to estimate continuous object parameters like human pose and object's viewpoint. In our experiment, we achieve state-of-the-art performance for view-classification on the PASCAL VOC'06 dataset.

Backprojection revisited: scalable multi-view object detection and similarity metrics for detections

Unsupervised domain adaptation aims to generalize a model from the label-rich source domain to the unlabeled target domain. Existing works mainly focus on aligning the global distribution statistics between source and target domains. However, they neglect distractions from the unexpected noisy samples in domain distribution estimation, leading to domain misalignment or even negative transfer. In this paper, we present an importance sampling method for domain adaptation (ISDA), to measure sample contributions according to their “informative” levels. In particular, informative samples, as well as outliers, can be effectively modeled using feature-norm and prediction entropy of the network. The importance of information is further formulated as the importance sampling losses in features and label spaces. In this way, the proposed model mitigates the noisy outliers while enhancing the important samples during domain alignment. In addition, our model is easy to implement yet effective, and it does not introduce any extra parameters. Extensive experiments on several benchmark datasets show that our method outperforms state-of-the-art methods under both the standard and partial domain adaptation settings.

Unsupervised Domain Adaptation via Importance Sampling

Recently Siamese-based trackers have shown their outstanding performance in visual object tracking community. But they seldom pay attention to the inter-branch interaction as well as intra-branch feature fusion from different convolution layers. In this paper, we build up a comprehensive Siamese network which consists of a mutual learning subnetwork (M-net) and a feature fusion subnetwork (F-net), to realize object tracking. Each of them is a Siamese network with special functions. M-net is designed to help the two branches mine the dependencies from each other, thus the object template is adaptively updated to a certain extent. F-net fuses different levels of convolutional features for full usage of spatial and semantic information. We also design a global-local channel attention (GLCA) module in F-net to capture the channel dependencies for a proper feature fusion. Our method takes ResNet as feature extractor and is trained offline in an end-to-end style. We evaluate our method in several famous benchmarks such as OTB2013, OTB2015, VOT2015, VOT2016, NFS and TC128. Extensive experimental results demonstrate our method achieves competitive results while maintaining a considerable real-time speed.

Mutual Learning and Feature Fusion Siamese Networks for Visual Object Tracking

Visual tracking is one of the fundamental problems in computer vision. Recently, some deep-learning-based tracking algorithms have been illustrating record-breaking performances. However, due to the high complexity of neural networks, most deep trackers suffer from low tracking speed and are, thus, impractical in many real-world applications. Some recently proposed deep trackers with smaller network structure achieve high efficiency while at the cost of significant decrease in precision. In this paper, we propose to transfer the deep feature, which is learned originally for image classification to the visual tracking domain. The domain adaptation is achieved via some “grafted” auxiliary networks, which are trained by regressing the object location in tracking frames. This adaptation improves the tracking performance significantly both on accuracy and efficiency. The yielded deep tracker is real time and also illustrates the state-of-the-art accuracies in the experiment involving two well-adopted benchmarks with more than 100 test videos. Furthermore, the adaptation is also naturally used for introducing the objectness concept into visual tracking. This removes a long-standing target ambiguity in visual tracking tasks, and we illustrate the empirical superiority of the more well-defined task.

Real-Time Deep Tracking via Corrective Domain Adaptation

Visual tracking is a fundamental problem in computer vision. However, due to the (sometimes) ambiguous target information given at the first frame, it has also been criticized as less well-posed compared with other tasks with clearly-defined targets, such as object detection and semantic segmentation. In this paper, we try to evaluate the importance of object category in visual tracking by tracking objects with known object types. The proposed algorithm, termed Deep-Track with Objectness (DTO), naturally combines the state-of-the-art deep-learning-based detectors and trackers, which essentially share a large part of the network. In DTO, a deep tracker, which is scale-fixed and sensitive to small translations tracks the object in a relative short lifespan. A deep detector, which is scale-changeable and robust to pose or illumination changes guides the deep tracker in a longer lifespan. As the deep tracker and detector share the main part of their networks, no much extra computation is imposed while the performance gain is significant. We test the proposed algorithm on two well-accepted benchmarks and on both of them, the proposed method increases the tracking accuracies remarkably compared with state-of-the-art visual trackers.

Deep tracking with objectness

We present a novel max-margin Hough transform with latent structure for joint object detection and pose estimation. Our method addresses the large appearance and shape variation of objects in multiple poses by integrating three key components: First, we propose a more robust appearance model by designing a patch dictionary with complementary features; In addition, we use a group of latent components to explicitly incorporate feature selection and pooling into the Hough-based object models; Furthermore, we adopt a multiple instance learning approach to handle the lack of correspondence among training instances with noisy bounding-box labels. We design a unified objective and an efficient approximate inference that alternates the search between object location and pose space. We demonstrate the efficacy of our approach by achieving the state-of-the-art performance on two detection and two joint estimation datasets.

Learning Hough Transform with Latent Structures for Joint Object Detection and Pose Estimation

Automatic evaluation of machine translations is an important task. Most existing evaluation metrics rely on matching the same word or letter n-grams. This strategy leads to poor results on Chinese translations because one has to rely merely on matching identical characters. In this article, we propose a new evaluation metric that allows different characters with the same or similar meaning to match. An Indirect Hidden Markov Model (IHMM) is proposed to align the Chinese translation with human references at the character level. In the model, the emission probabilities are estimated by character similarity, including character semantic similarity and character surface similarity, and transition probabilities are estimated by a heuristic distance-based distortion model. When evaluating the submitted output of English-to-Chinese translation systems in the IWSLT’08 CT-EC and NIST’08 EC tasks, the experimental results indicate that the proposed metric has a significantly better correlation with human evaluation than the state-of-the-art machine translation metrics (i.e., BLEU, Meteor Universal, and TESLA-CELAB). This study shows that it is important to allow different characters to match in the evaluation of Chinese translations and that the IHMM is a reasonable approach for the alignment of Chinese characters.

Modeling Monolingual Character Alignment for Automatic Evaluation of Chinese Translation

Heterogeneous gap among different modalities emerges as one of the critical issues in multimedia retrieval areas. Unlike traditional unimodal cases, where raw features are extracted and directly measured, the heterogeneous nature of crossmodal tasks requires the intrinsic semantic representation to be compared in a unified framework. Based on a flexible “feature up-lifting and down projecting” mechanism, this paper studies the learning of crossmodal semantic features that can be retrieved across different modalities. Two effective methods are proposed to mine semantic correlations. One is for traditional handcrafted features, and the other is based on deep neural network. We treat them respectively as normal and deep version of our proposed shared discriminative semantic representation learning (SDSRL) framework. We evaluate both of these two methods on two public multimodal datasets for crossmodal and unimodal retrieval tasks. The experimental results demonstrate that our proposed methods outperform the compared baselines and achieve state-of-the-art performance in most scenarios.

Mingwen Wang

Papers

Real-Time Deep Tracking via Corrective Domain Adaptation

Deep tracking with objectness

Learning Hough Transform with Latent Structures for Joint Object Detection and Pose Estimation

Modeling Monolingual Character Alignment for Automatic Evaluation of Chinese Translation

Learning Discriminative Representations for Semantic Cross Media Retrieval