Video caption refers to generating a descriptive sentence for a specific short video clip automatically, which has achieved remarkable success recently. However, most of the existing methods focus more on visual information while ignoring the synchronized audio cues. We propose three multimodal deep fusion strategies to maximize the benefits of visual-audio resonance information. The first one explores the impact on cross-modalities feature fusion from low to high order. The second establishes the visual-audio short-term dependency by sharing weights of corresponding front-end networks. The third extends the temporal dependency to long-term through sharing multimodal memory across visual and audio modalities. Extensive experiments have validated the effectiveness of our three cross-modalities fusion strategies on two benchmark datasets, including Microsoft Research Video to Text (MSRVTT) and Microsoft Video Description (MSVD). It is worth mentioning that sharing weight can coordinate visual- audio feature fusion effectively and achieve the state-of-art performance on both BELU and METEOR metrics. Furthermore, we first propose a dynamic multimodal feature fusion framework to deal with the part modalities missing case. Experimental results demonstrate that even in the audio absence mode, we can still obtain comparable results with the aid of the additional audio modality inference module.

Face Parsing assigns every pixel in a facial image with a semantic label, which could be applied in various applications including face recognition, facial beautification, affective computing and animation. While lots of progress have been made in this field, current state-of-the-art methods still fail to extract real effective feature and restore accurate score map, especially for those facial parts which have large variations of deformation and fairly similar appearance, e.g. mouth, eyes and thin eyebrows. In this paper, we propose a novel pixel-wise face parsing method called Residual Encoder Decoder Network (RED-Net), which combines a feature-rich encoder-decoder framework with adaptive prior mechanism. Our encoder-decoder framework extracts feature with ResNet and decodes the feature by elaborately fusing the residual architectures in to deconvolution. This framework learns more effective feature comparing to that learnt by decoding with interpolation or classic deconvolution operations. To overcome the appearance ambiguity between facial parts, an adaptive prior mechanism is proposed in term of the decoder prediction confidence, allowing refining the final result. The experimental results on two public datasets demonstrate that our method outperforms the state-of-the-arts significantly, achieving improvements of F-measure from 0.854 to 0.905 on Helen dataset, and pixel accuracy from 95.12% to 97.59% on the LFW dataset. In particular, convincing qualitative examples show that our method parses eye, eyebrow, and lip regins more accurately.

Vehicle re-identification (re-ID) is to identify the same vehicle across different cameras. It’s a significant but challenging topic, which has received little attention due to the complex intra-class and inter-class variation of vehicle images and the lack of large-scale vehicle re-ID dataset. Previous methods focus on pulling images from different vehicles apart but neglect the discrimination between vehicles from different vehicle models, which is actually quite important to obtain a correct ranking order for vehicle re-ID. In this paper, we learn a structured feature embedding for vehicle re-ID with a novel coarse-to-fine ranking loss to pull images of the same vehicle as close as possible and achieve discrimination between images from different vehicles as well as vehicles from different vehicle models. In the learnt feature space, both intra-class compactness and inter-class distinction are well guaranteed and the Euclidean distance between features directly reflects the semantic similarity of vehicle images. Furthermore, we build so far the largest vehicle re-ID dataset "Vehicle-1M," which involves nearly 1 million images captured in various surveillance scenarios. Experimental results on "Vehicle-1M" and "VehicleID" demonstrate the superiority of our proposed approach.

Continuous Sign Language Translation (SLT) is a challenging task due to its specific linguistics under sequential gesture variation without word alignment. Current hybrid HMM and CTC (Connectionist temporal classification) based models are proposed to solve frame or word level alignment. They may fail to tackle the cases with messing word order corresponding to visual content in sentences. To solve the issue, this paper proposes a hierarchical-LSTM (HLSTM) encoder-decoder model with visual content and word embedding for SLT. It tackles different granularities by conveying spatio-temporal transitions among frames, clips and viseme units. It firstly explores spatio-temporal cues of video clips by 3D CNN and packs appropriate visemes by online key clip mining with adaptive variable-length. After pooling on recurrent outputs of the top layer of HLSTM, a temporal attention-aware weighting mechanism is proposed to balance the intrinsic relationship among viseme source positions. At last, another two LSTM layers are used to separately recurse viseme vectors and translate semantic. After preserving original visual content by 3D CNN and the top layer of HLSTM, it shortens the encoding time step of the bottom two LSTM layers with less computational complexity while attaining more nonlinearity. Our proposed model exhibits promising performance on singer-independent test with seen sentences and also outperforms the comparison algorithms on unseen sentences.

The existing image captioning approaches typically train a one-stage sentence decoder, which is difficult to generate rich fine-grained descriptions. On the other hand, multi-stage image caption model is hard to train due to the vanishing gradient problem. In this paper, we propose a coarse-to-fine multi-stage prediction framework for image captioning, composed of multiple decoders each of which operates on the output of the previous stage, producing increasingly refined image descriptions. Our proposed learning approach addresses the difficulty of vanishing gradients during training by providing a learning objective function that enforces intermediate supervisions. Particularly, we optimize our model with a reinforcement learning approach which utilizes the output of each intermediate decoder's test-time inference algorithm as well as the output of its preceding decoder to normalize the rewards, which simultaneously solves the well-known exposure bias problem and the loss-evaluation mismatch problem. We extensively evaluate the proposed approach on MSCOCO and show that our approach can achieve the state-of-the-art performance.

Deep neural network (DNN) architecture based models have high expressive power and learning capacity. However, they are essentially a black box method since it is not easy to mathematically formulate the functions that are learned within its many layers of representation. Realizing this, many researchers have started to design methods to exploit the drawbacks of deep learning based algorithms questioning their robustness and exposing their singularities. In this paper, we attempt to unravel three aspects related to the robustness of DNNs for face recognition: (i) assessing the impact of deep architectures for face recognition in terms of vulnerabilities to attacks inspired by commonly observed distortions in the real world that are well handled by shallow learning methods along with learning based adversaries; (ii) detecting the singularities by characterizing abnormal filter response behavior in the hidden layers of deep networks; and (iii) making corrections to the processing pipeline to alleviate the problem. Our experimental evaluation using multiple open-source DNN-based face recognition networks, including OpenFace and VGG-Face, and two publicly available databases (MEDS and PaSC) demonstrates that the performance of deep learning based face recognition algorithms can suffer greatly in the presence of such distortions. The proposed method is also compared with existing detection algorithms and the results show that it is able to detect the attacks with very high accuracy by suitably designing a classifier using the response of the hidden layers in the network. Finally, we present several effective countermeasures to mitigate the impact of adversarial attacks and improve the overall robustness of DNN-based face recognition.

In this paper, we propose a pose grammar to tackle the problem of 3D human pose estimation. Our model directly takes 2D pose as input and learns a generalized 2D-3D mapping function. The proposed model consists of a base network which efficiently captures pose-aligned features and a hierarchy of Bi-directional RNNs (BRNN) on the top to explicitly incorporate a set of knowledge regarding human body configuration (i.e., kinematics, symmetry, motor coordination). The proposed model thus enforces high-level constraints over human poses. In learning, we develop a pose sample simulator to augment training samples in virtual camera views, which further improves our model generalizability. We validate our method on public 3D human pose benchmarks and propose a new evaluation protocol working on cross-view setting to verify the generalization capability of different methods. We empirically observe that most state-of-the-art methods encounter difficulty under such setting while our method can well handle such challenges.

Feature extraction is a critical step in the task of action recognition. Hand-crafted features are often restricted because of their fixed forms and deep learning features are more effective but need large-scale labeled data for training. In this paper, we propose a new hierarchical Nonlinear Orthogonal Adaptive-Subspace Self-Organizing Map(NOASSOM) to adaptively and learn effective features from data without supervision. NOASSOM is extended from Adaptive-Subspace Self-Organizing Map (ASSOM) which only deals with linear data and is trained with supervision by the labeled data. Firstly, by adding a nonlinear orthogonal map layer, NOASSOM is able to handle the nonlinear input data and it avoids defining the specific form of the nonlinear orthogonal map by a kernel trick. Secondly, we modify loss function of ASSOM such that every input sample is used to train model individually. In this way, NOASSOM effectively learns the statistic patterns from data without supervision. Thirdly, we propose a hierarchical NOASSOM to extract more representative features. Finally, we apply the proposed hierarchical NOASSOM to efficiently describe the appearance and motion information around trajectories for action recognition. Experimental results on widely used datasets show that our method has superior performance than many state-of-the-art hand-crafted features and deep learning features based methods.

In recent years considerable research efforts have been devoted to compression techniques of convolutional neural networks (CNNs). Many works so far have focused on CNN connection pruning methods which produce sparse parameter tensors in convolutional or fully-connected layers. It has been demonstrated in several studies that even simple methods can effectively eliminate connections of a CNN. However, since these methods make parameter tensors just sparser but no smaller, the compression may not transfer directly to acceleration without support from specially designed hardware. In this paper, we propose an iterative approach named Auto-balanced Filter Pruning, where we pre-train the network in an innovative auto-balanced way to transfer the representational capacity of its convolutional layers to a fraction of the filters, prune the redundant ones, then re-train it to restore the accuracy. In this way, a smaller version of the original network is learned and the floating-point operations (FLOPs) are reduced. By applying this method on several common CNNs, we show that a large portion of the filters can be discarded without obvious accuracy drop, leading to significant reduction of computational burdens. Concretely, we reduce the inference cost of LeNet-5 on MNIST, VGG-16 and ResNet-56 on CIFAR-10 by 95.1%, 79.7% and 60.9%, respectively.

Most state-of-the-art scene text detection algorithms are deep learning based methods that depend on bounding box regression and perform at least two kinds of predictions: text/non-text classification and location regression. Regression plays a key role in the acquisition of bounding boxes in these methods, but it is not indispensable because text/non-text prediction can also be considered as a kind of semantic segmentation that contains full location information in itself. However, text instances in scene images often lie very close to each other, making them very difficult to separate via semantic segmentation. Therefore, instance segmentation is needed to address this problem. In this paper, PixelLink, a novel scene text detection algorithm based on instance segmentation, is proposed. Text instances are first segmented out by linking pixels within the same instance together. Text bounding boxes are then extracted directly from the segmentation result without location regression. Experiments show that, compared with regression-based methods, PixelLink can achieve better or comparable performance on several benchmarks, while requiring many fewer training iterations and less training data.