Research papers and code for "Wenjun Zeng":
Recently, Siamese network based trackers have received tremendous interest for their fast tracking speed and high performance. Despite the great success, this tracking framework still suffers from several limitations. First, it cannot properly handle large object rotation. Second, tracking gets easily distracted when the background contains salient objects. In this paper, we propose two simple yet effective mechanisms, namely angle estimation and spatial masking, to address these issues. The objective is to extract more representative features so that a better match can be obtained between the same object from different frames. The resulting tracker, named Siam-BM, not only significantly improves the tracking performance, but more importantly maintains the realtime capability. Evaluations on the VOT2017 dataset show that Siam-BM achieves an EAO of 0.335, which makes it the best-performing realtime tracker to date.

* This paper is accepted by ECCV Visual Object Tracking Challenge Workshop VOT2018
Click to Read Paper and Get Code
Observing that Semantic features learned in an image classification task and Appearance features learned in a similarity matching task complement each other, we build a twofold Siamese network, named SA-Siam, for real-time object tracking. SA-Siam is composed of a semantic branch and an appearance branch. Each branch is a similarity-learning Siamese network. An important design choice in SA-Siam is to separately train the two branches to keep the heterogeneity of the two types of features. In addition, we propose a channel attention mechanism for the semantic branch. Channel-wise weights are computed according to the channel activations around the target position. While the inherited architecture from SiamFC \cite{SiamFC} allows our tracker to operate beyond real-time, the twofold design and the attention mechanism significantly improve the tracking performance. The proposed SA-Siam outperforms all other real-time trackers by a large margin on OTB-2013/50/100 benchmarks.

* Accepted by CVPR'18
Click to Read Paper and Get Code
The past decade has witnessed great success in applying deep learning to enhance the quality of compressed video. However, the existing approaches aim at quality enhancement on a single frame, or only using fixed neighboring frames. Thus they fail to take full advantage of the inter-frame correlation in the video. This paper proposes the Quality-Gated Convolutional Long Short-Term Memory (QG-ConvLSTM) network with bi-directional recurrent structure to fully exploit the advantageous information in a large range of frames. More importantly, due to the obvious quality fluctuation among compressed frames, higher quality frames can provide more useful information for other frames to enhance quality. Therefore, we propose learning the "forget" and "input" gates in the ConvLSTM cell from quality-related features. As such, the frames with various quality contribute to the memory in ConvLSTM with different importance, making the information of each frame reasonably and adequately used. Finally, the experiments validate the effectiveness of our QG-ConvLSTM approach in advancing the state-of-the-art quality enhancement of compressed video, and the ablation study shows that our QG-ConvLSTM approach is learnt to make a trade-off between quality and correlation when leveraging multi-frame information. The project page: https://github.com/ryangchn/QG-ConvLSTM.git.

* Accepted to IEEE International Conference on Multimedia and Expo (ICME) 2019
Click to Read Paper and Get Code
The greatest challenge facing visual object tracking is the simultaneous requirements on robustness and discrimination power. In this paper, we propose a SiamFC-based tracker, named SPM-Tracker, to tackle this challenge. The basic idea is to address the two requirements in two separate matching stages. Robustness is strengthened in the coarse matching (CM) stage through generalized training while discrimination power is enhanced in the fine matching (FM) stage through a distance learning network. The two stages are connected in series as the input proposals of the FM stage are generated by the CM stage. They are also connected in parallel as the matching scores and box location refinements are fused to generate the final results. This innovative series-parallel structure takes advantage of both stages and results in superior performance. The proposed SPM-Tracker, running at 120fps on GPU, achieves an AUC of 0.687 on OTB-100 and an EAO of 0.434 on VOT-16, exceeding other real-time trackers by a notable margin.

* to appear in CVPR'19
Click to Read Paper and Get Code
The recent success of deep networks has significantly advanced 3D human pose estimation from 2D images. The diversity of capturing viewpoints and the flexibility of the human poses, however, remain some significant challenges. In this paper, we propose a view invariant 3D human pose estimation module to alleviate the effects of viewpoint diversity. The framework consists of a base network, which provides an initial estimation of a 3D pose, a view-invariant hierarchical correction network (VI-HC) on top of that to learn the 3D pose refinement under consistent views, and a view-invariant discriminative network (VID) to enforce high-level constraints over body configurations. In VI-HC, the initial 3D pose inputs are automatically transformed to consistent views for further refinements at the global body and local body parts level, respectively. For the VID, under consistent viewpoints, we use adversarial learning to differentiate between estimated poses and real poses to avoid implausible 3D poses. Experimental results demonstrate that the consistent viewpoints can dramatically enhance the performance. Our module shows robustness for different 3D pose base networks and achieves a significant improvement (about 9%) over a powerful baseline on the public 3D pose estimation benchmark Human3.6M.

Click to Read Paper and Get Code
We propose a densely semantically aligned person re-identification (re-ID) framework. It fundamentally addresses the body misalignment problem caused by pose/viewpoint variations, imperfect person detection, occlusion,etc.. By leveraging the estimation of the dense semantics of a person image, we construct a set of densely semantically aligned part images (DSAP-images), where the same spatial positions have the same semantics across different person images. We design a two-stream network that consists of a main full image stream (MF-Stream) and a densely semantically-aligned guiding stream (DSAG-Stream). The DSAG-Stream, with the DSAP-images as input, acts as a regulator to guide the MF-Stream to learn densely semantically aligned features from the original image. In the inference, the DSAG-Stream is discarded and only the MF-Stream is needed, which makes the inference system computationally efficient and robust. To our best knowledge, we are the first to make use of fine grained semantics for addressing misalignment problems for re-ID. Our method achieves rank-1 accuracy of 78.9% (new protocol) on the CUHK03 dataset, 90.4% on the CUHK01 dataset, and 95.7% on the Market1501 dataset, outperforming state-of-the-art methods.

Click to Read Paper and Get Code
State-of-the-art object detectors and trackers are developing fast. Trackers are in general more efficient than detectors but bear the risk of drifting. A question is hence raised -- how to improve the accuracy of video object detection/tracking by utilizing the existing detectors and trackers within a given time budget? A baseline is frame skipping -- detecting every N-th frames and tracking for the frames in between. This baseline, however, is suboptimal since the detection frequency should depend on the tracking quality. To this end, we propose a scheduler network, which determines to detect or track at a certain frame, as a generalization of Siamese trackers. Although being light-weight and simple in structure, the scheduler network is more effective than the frame skipping baselines and flow-based approaches, as validated on ImageNet VID dataset in video object detection/tracking.

* Accepted to AAAI 2019
Click to Read Paper and Get Code
Model update lies at the heart of object tracking.Generally, model update is formulated as an online learning problem where a target model is learned over the online training dataset. Our key innovation is to \emph{learn the online learning algorithm itself using large number of offline videos}, i.e., \emph{learning to update}. The learned updater takes as input the online training dataset and outputs an updated target model. As a first attempt, we design the learned updater based on recurrent neural networks (RNNs) and demonstrate its application in a template-based tracker and a correlation filter-based tracker. Our learned updater consistently improves the base trackers and runs faster than realtime on GPU while requiring small memory footprint during testing. Experiments on standard benchmarks demonstrate that our learned updater outperforms commonly used update baselines including the efficient exponential moving average (EMA)-based update and the well-designed stochastic gradient descent (SGD)-based update. Equipped with our learned updater, the template-based tracker achieves state-of-the-art performance among realtime trackers on GPU.

Click to Read Paper and Get Code
With the rapid development of social network and multimedia technology, customized image and video stylization has been widely used for various social-media applications. In this paper, we explore the problem of exemplar-based photo style transfer, which provides a flexible and convenient way to invoke fantastic visual impression. Rather than investigating some fixed artistic patterns to represent certain styles as was done in some previous works, our work emphasizes styles related to a series of visual effects in the photograph, e.g. color, tone, and contrast. We propose a photo stylistic brush, an automatic robust style transfer approach based on Superpixel-based BIpartite Graph (SuperBIG). A two-step bipartite graph algorithm with different granularity levels is employed to aggregate pixels into superpixels and find their correspondences. In the first step, with the extracted hierarchical features, a bipartite graph is constructed to describe the content similarity for pixel partition to produce superpixels. In the second step, superpixels in the input/reference image are rematched to form a new superpixel-based bipartite graph, and superpixel-level correspondences are generated by a bipartite matching. Finally, the refined correspondence guides SuperBIG to perform the transformation in a decorrelated color space. Extensive experimental results demonstrate the effectiveness and robustness of the proposed method for transferring various styles of exemplar images, even for some challenging cases, such as night images.

Click to Read Paper and Get Code
In this paper, we present a novel deep learning approach, deeply-fused nets. The central idea of our approach is deep fusion, i.e., combine the intermediate representations of base networks, where the fused output serves as the input of the remaining part of each base network, and perform such combinations deeply over several intermediate representations. The resulting deeply fused net enjoys several benefits. First, it is able to learn multi-scale representations as it enjoys the benefits of more base networks, which could form the same fused network, other than the initial group of base networks. Second, in our suggested fused net formed by one deep and one shallow base networks, the flows of the information from the earlier intermediate layer of the deep base network to the output and from the input to the later intermediate layer of the deep base network are both improved. Last, the deep and shallow base networks are jointly learnt and can benefit from each other. More interestingly, the essential depth of a fused net composed from a deep base network and a shallow base network is reduced because the fused net could be composed from a less deep base network, and thus training the fused net is less difficult than training the initial deep base network. Empirical results demonstrate that our approach achieves superior performance over two closely-related methods, ResNet and Highway, and competitive performance compared to the state-of-the-arts.

Click to Read Paper and Get Code
Person re-identification (reID) aims to match person images to retrieve the ones with the same identity. This is a challenging task, as the images to be matched are generally semantically misaligned due to the diversity of human poses and capture viewpoints, incompleteness of the visible bodies (due to occlusion), etc. In this paper, we propose a framework that drives the reID network to learn semantics-aligned feature representation through delicate supervision designs. Specifically, we build a Semantics Aligning Network (SAN) which consists of a base network as encoder (SA-Enc) for re-ID, and a decoder (SA-Dec) for reconstructing/regressing the densely semantics aligned full texture image. We jointly train the SAN under the supervisions of person re-identification and aligned texture generation. Moreover, at the decoder, besides the reconstruction loss, we add triplet reID constraints/losses over the feature maps as the perceptual losses. The decoder is discarded in the inference/test and thus our scheme is computationally efficient. Ablation studies demonstrate the effectiveness of our design. We achieve the state-of-the-art performances on the benchmark datasets CUHK03, Market1501, MSMT17, and the partial person reID dataset Partial REID.

Click to Read Paper and Get Code
Objects in an image exhibit diverse scales. Adaptive receptive fields are expected to catch suitable range of context for accurate pixel level semantic prediction for handling objects of diverse sizes. Recently, atrous convolution with different dilation rates has been used to generate features of multi-scales through several branches and these features are fused for prediction. However, there is a lack of explicit interaction among the branches to adaptively make full use of the contexts. In this paper, we propose a Content-Adaptive Scale Interaction Network (CaseNet) to exploit the multi-scale features for scene parsing. We build the CaseNet based on the classic Atrous Spatial Pyramid Pooling (ASPP) module, followed by the proposed contextual scale interaction (CSI) module, and the scale adaptation (SA) module. Specifically, first, for each spatial position, we enable context interaction among different scales through scale-aware non-local operations across the scales, \ie, CSI module, which facilitates the generation of flexible mixed receptive fields, instead of a traditional flat one. Second, the scale adaptation module (SA) explicitly and softly selects the suitable scale for each spatial position and each channel. Ablation studies demonstrate the effectiveness of the proposed modules. We achieve state-of-the-art performance on three scene parsing benchmarks Cityscapes, ADE20K and LIP.

Click to Read Paper and Get Code
Attention mechanism aims to increase the representation power by focusing on important features and suppressing unnecessary ones. For convolutional neural networks (CNNs), attention is typically learned with local convolutions, which ignores the global information and the hidden relation. How to efficiently exploit the long-range context to globally learn attention is underexplored. In this paper, we propose an effective Relation-Aware Global Attention (RGA) module for CNNs to fully exploit the global correlations to infer the attention. Specifically, when computing the attention at a feature position, in order to grasp information of global scope, we propose to stack the relations, i.e., its pairwise correlations/affinities with all the feature positions, and the feature itself together for learning the attention with convolutional operations. Given an intermediate feature map, we have validated the effectiveness of this design across both the spatial and channel dimensions. When applied to the task of person re-identification, our model achieves the state-of-the-art performance. Extensive ablation studies demonstrate that our RGA can significantly enhance the feature representation power. We further demonstrate the general applicability of RGA to vision tasks by applying it to the scene segmentation and image classification tasks resulting in consistent performance improvement.

Click to Read Paper and Get Code
Skeleton-based human action recognition has attracted a lot of interests. Recently, there is a trend of using deep feedforward neural networks to model the skeleton sequence which takes the 2D spatio-temporal map derived from the 3D coordinates of joints as input. Some semantics of the joints (frame index and joint type) are implicitly captured and exploited by large receptive fields of deep convolutions at the cost of high complexity. In this paper, we propose a simple yet effective semantics-guided neural network (SGN) for skeleton-based action recognition. We explicitly introduce the high level semantics of joints as part of the network input to enhance the feature representation capability. The model exploits the global and local information through two semantics-aware graph convolutional layers followed by a convolutional layer. We first leverage the semantics and dynamics (coordinate and velocity) of joints to learn a content adaptive graph for capturing the global spatio-temporal correlations of joints. Then a convolutional layer is used to further enhance the representation power of the features. With an order of magnitude smaller model size and higher speed than some previous works, SGN achieves the state-of-the-art performance on the NTU, SYSU, and N-UCLA datasets. Experimental results demonstrate the effectiveness of explicitly exploiting semantic information in reducing model complexity and improving the recognition accuracy.

Click to Read Paper and Get Code
Compared with object detection in static images, object detection in videos is more challenging due to degraded image qualities. An effective way to address this problem is to exploit temporal contexts by linking the same object across video to form tubelets and aggregating classification scores in the tubelets. In this paper, we focus on obtaining high quality object linking results for better classification. Unlike previous methods that link objects by checking boxes between neighboring frames, we propose to link in the same frame. To achieve this goal, we extend prior methods in following aspects: (1) a cuboid proposal network that extracts spatio-temporal candidate cuboids which bound the movement of objects; (2) a short tubelet detection network that detects short tubelets in short video segments; (3) a short tubelet linking algorithm that links temporally-overlapping short tubelets to form long tubelets. Experiments on the ImageNet VID dataset show that our method outperforms both the static image detector and the previous state of the art. In particular, our method improves results by 8.8% over the static image detector for fast moving objects.

Click to Read Paper and Get Code
A deep residual network, built by stacking a sequence of residual blocks, is easy to train, because identity mappings skip residual branches and thus improve information flow. To further reduce the training difficulty, we present a simple network architecture, deep merge-and-run neural networks. The novelty lies in a modularized building block, merge-and-run block, which assembles residual branches in parallel through a merge-and-run mapping: Average the inputs of these residual branches (Merge), and add the average to the output of each residual branch as the input of the subsequent residual branch (Run), respectively. We show that the merge-and-run mapping is a linear idempotent function in which the transformation matrix is idempotent, and thus improves information flow, making training easy. In comparison to residual networks, our networks enjoy compelling advantages: they contain much shorter paths, and the width, i.e., the number of channels, is increased. We evaluate the performance on the standard recognition tasks. Our approach demonstrates consistent improvements over ResNets with the comparable setup, and achieves competitive results (e.g., $3.57\%$ testing error on CIFAR-$10$, $19.00\%$ on CIFAR-$100$, $1.51\%$ on SVHN).

Click to Read Paper and Get Code
Human action recognition is an important task in computer vision. Extracting discriminative spatial and temporal features to model the spatial and temporal evolutions of different actions plays a key role in accomplishing this task. In this work, we propose an end-to-end spatial and temporal attention model for human action recognition from skeleton data. We build our model on top of the Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM), which learns to selectively focus on discriminative joints of skeleton within each frame of the inputs and pays different levels of attention to the outputs of different frames. Furthermore, to ensure effective training of the network, we propose a regularized cross-entropy loss to drive the model learning process and develop a joint training strategy accordingly. Experimental results demonstrate the effectiveness of the proposed model,both on the small human action recognition data set of SBU and the currently largest NTU dataset.

Click to Read Paper and Get Code
The emergence of one-shot approaches has greatly advanced the research on neural architecture search (NAS). Recent approaches train an over-parameterized super-network (one-shot model) and then sample and evaluate a number of sub-networks, which inherit weights from the one-shot model. The overall searching cost is significantly reduced as training is avoided for sub-networks. However, the network sampling process is casually treated and the inherited weights from an independently trained super-network perform sub-optimally for sub-networks. In this paper, we propose a novel one-shot NAS scheme to address the above issues. The key innovation is to explicitly estimate the joint a posteriori distribution over network architecture and weights, and sample networks for evaluation according to it. This brings two benefits. First, network sampling under the guidance of a posteriori probability is more efficient than conventional random or uniform sampling. Second, the network architecture and its weights are sampled as a pair to alleviate the sub-optimal weights problem. Note that estimating the joint a posteriori distribution is not a trivial problem. By adopting variational methods and introducing a hybrid network representation, we convert the distribution approximation problem into an end-to-end neural network training problem which is neatly approached by variational dropout. As a result, the proposed method reduces the number of sampled sub-networks by orders of magnitude. We validate our method on the fundamental image classification task. Results on Cifar-10, Cifar-100 and ImageNet show that our method strikes the best trade-off between precision and speed among NAS methods. On Cifar-10, we speed up the searching process by 20x and achieve a higher precision than the best network found by existing NAS methods.

Click to Read Paper and Get Code
Scene graph construction / visual relationship detection from an image aims to give a precise structural description of the objects (nodes) and their relationships (edges). The mutual promotion of object detection and relationship detection is important for enhancing their individual performance. In this work, we propose a new framework, called semantics guided graph relation neural network (SGRN), for effective visual relationship detection. First, to boost the object detection accuracy, we introduce a source-target class cognoscitive transformation that transforms the features of the co-occurent objects to the target object domain to refine the visual features. Similarly, source-target cognoscitive transformations are used to refine features of objects from features of relations, and vice versa. Second, to boost the relation detection accuracy, besides the visual features of the paired objects, we embed the class probability of the object and subject separately to provide high level semantic information. In addition, to reduce the search space of relationships, we design a semantics-aware relationship filter to exclude those object pairs that have no relation. We evaluate our approach on the Visual Genome dataset and it achieves the state-of-the-art performance for visual relationship detection. Additionally, Our approach also significantly improves the object detection performance (i.e. 4.2\% in mAP accuracy).

Click to Read Paper and Get Code
Deep learning models have enjoyed great success for image related computer vision tasks like image classification and object detection. For video related tasks like human action recognition, however, the advancements are not as significant yet. The main challenge is the lack of effective and efficient models in modeling the rich temporal spatial information in a video. We introduce a simple yet effective operation, termed Temporal-Spatial Mapping (TSM), for capturing the temporal evolution of the frames by jointly analyzing all the frames of a video. We propose a video level 2D feature representation by transforming the convolutional features of all frames to a 2D feature map, referred to as VideoMap. With each row being the vectorized feature representation of a frame, the temporal-spatial features are compactly represented, while the temporal dynamic evolution is also well embedded. Based on the VideoMap representation, we further propose a temporal attention model within a shallow convolutional neural network to efficiently exploit the temporal-spatial dynamics. The experiment results show that the proposed scheme achieves the state-of-the-art performance, with 4.2% accuracy gain over Temporal Segment Network (TSN), a competing baseline method, on the challenging human action benchmark dataset HMDB51.

Click to Read Paper and Get Code