Adversarial examples are commonly viewed as a threat to ConvNets. Here we present an opposite perspective: adversarial examples can be used to improve image recognition models if harnessed in the right manner. We propose AdvProp, an enhanced adversarial training scheme which treats adversarial examples as additional examples, to prevent overfitting. Key to our method is the usage of a separate auxiliary batch norm for adversarial examples, as they have different underlying distributions to normal examples. We show that AdvProp improves a wide range of models on various image recognition tasks and performs better when the models are bigger. For instance, by applying AdvProp to the latest EfficientNet-B7 [28] on ImageNet, we achieve significant improvements on ImageNet (+0.7%), ImageNet-C (+6.5%), ImageNet-A (+7.0%), Stylized-ImageNet (+4.8%). With an enhanced EfficientNet-B8, our method achieves the state-of-the-art 85.5% ImageNet top-1 accuracy without extra data. This result even surpasses the best model in [20] which is trained with 3.5B Instagram images (~3000X more than ImageNet) and ~9.4X more parameters. Code and models will be made publicly available.

The labeling cost of large number of bounding boxes is one of the main challenges for training modern object detectors. To reduce the dependence on expensive bounding box annotations, we propose a new semi-supervised object detection formulation, in which a few seed box level annotations and a large scale of image level annotations are used to train the detector. We adopt a training-mining framework, which is widely used in weakly supervised object detection tasks. However, the mining process inherently introduces various kinds of labelling noises: false negatives, false positives and inaccurate boundaries, which can be harmful for training the standard object detectors (e.g. Faster RCNN). We propose a novel NOise Tolerant Ensemble RCNN (NOTE-RCNN) object detector to handle such noisy labels. Comparing to standard Faster RCNN, it contains three highlights: an ensemble of two classification heads and a distillation head to avoid overfitting on noisy labels and improve the mining precision, masking the negative sample loss in box predictor to avoid the harm of false negative labels, and training box regression head only on seed annotations to eliminate the harm from inaccurate boundaries of mined bounding boxes. We evaluate the methods on ILSVRC 2013 and MSCOCO 2017 dataset; we observe that the detection accuracy consistently improves as we iterate between mining and training steps, and state-of-the-art performance is achieved.

Modeling the relationship among human joints is one of the most important components in human pose estimation. Most of previous methods define this relationship as a geometric constraint on the relative locations of two neighboring joints. In this constraint, the local appearance of the region connecting two neighboring joints is ignored. However, discarding this image appearance leads to some severe problems, such as double-counting and localization failure when the human pose is rare in the training dataset. Moreover, this image appearance, called human limb, plays an important role in human pose estimation in human visual system. Due to these reasons, we propose to solve a new task: human limb detection, which aims at detecting and representing this local image appearance. We combine this task with human joint localization as a unified framework. After getting the initial detections, we design a two-steps graphical model to capture the spatial relationship among human joints and limbs in a coarse to fine way. We evaluate the proposed method on two widely used datasets for human pose estimation: 1) frame labeled in cinema and 2) leeds sports pose datasets. The experiments results show the effectiveness of our method.

Quantization methods are crucial for efficient nearest neighbor search in many applications such as image, music, or product search. As mobile devices are becoming increasingly more popular, the quantization methods on mobile devices are more important, because a large portion of the search queries are becoming performed on mobile devices. One important characteristic of the communication on mobile devices is the inherent unreliability of their communication channels. In order to adapt the quality changes of the communication channels, we need to change the coding length of the quantization accordingly. The existing quantization methods use fixed-length codebooks, and it is expensive to retrain another codebook with different coding length. In this paper, we propose a novel variable length product quantization framework that consists of a set of fast universal scalar quantizers. The framework is capable of producing variable length quantization without retraining the codebook. Each data vector is transformed into a new space to reduce the correlation across dimensions. A proper number of bits is allocated to represent the scalar component in each dimension according to the given coding length. For each component, we estimate its probability density function (PDF) and design an efficient universal scalar quantizer based on the PDF and the allocated bits. To reduce distortion, we learn a Gaussian mixture model for the data. The experimental results show that, compared to state-of-the-art product quantization methods, our approach can construct the codebooks online for variable coding lengths and achieve the comparable performance.

Incorporating multi-scale features in fully convolutional neural networks (FCNs) has been a key element to achieving state-of-the-art performance on semantic image segmentation. One common way to extract multi-scale features is to feed multiple resized input images to a shared deep network and then merge the resulting features for pixelwise classification. In this work, we propose an attention mechanism that learns to softly weight the multi-scale features at each pixel location. We adapt a state-of-the-art semantic image segmentation model, which we jointly train with multi-scale input images and the attention model. The proposed attention model not only outperforms averageand max-pooling, but allows us to diagnostically visualize the importance of features at different positions and scales. Moreover, we show that adding extra supervision to the output at each scale is essential to achieving excellent performance when merging multi-scale features. We demonstrate the effectiveness of our model with extensive experiments on three challenging datasets, including PASCAL-Person-Part, PASCAL VOC 2012 and a subset of MS-COCO 2014.

While deep convolutional neural networks (CNNs) have shown a great success in single-label image classification, it is important to note that real world images generally contain multiple labels, which could correspond to different objects, scenes, actions and attributes in an image. Traditional approaches to multi-label image classification learn independent classifiers for each category and employ ranking or thresholding on the classification results. These techniques, although working well, fail to explicitly exploit the label dependencies in an image. In this paper, we utilize recurrent neural networks (RNNs) to address this problem. Combined with CNNs, the proposed CNN-RNN framework learns a joint image-label embedding to characterize the semantic label dependency as well as the image-label relevance, and it can be trained end-to-end from scratch to integrate both information in a unified framework. Experimental results on public benchmark datasets demonstrate that the proposed architecture achieves better performance than the state-of-the-art multi-label classification models.

We present an approach that exploits hierarchical Recurrent Neural Networks (RNNs) to tackle the video captioning problem, i.e., generating one or multiple sentences to describe a realistic video. Our hierarchical framework contains a sentence generator and a paragraph generator. The sentence generator produces one simple short sentence that describes a specific short video interval. It exploits both temporal-and spatial-attention mechanisms to selectively focus on visual elements during generation. The paragraph generator captures the inter-sentence dependency by taking as input the sentential embedding produced by the sentence generator, combining it with the paragraph history, and outputting the new initial state for the sentence generator. We evaluate our approach on two large-scale benchmark datasets: YouTubeClips and TACoS-MultiLevel. The experiments demonstrate that our approach significantly outperforms the current state-of-the-art methods with BLEU@4 scores 0.499 and 0.305 respectively.

Modeling the relationship among human joints is one of the most important components in human pose estimation. Previous methods usually define this relationship as geometric constraints on the relative location of two neighboring joints. In this definition, the local image appearance of the region connecting two neighboring joints is ignored. In fact, this image appearance, called human limb, plays an important role in human joint localization in human visual system. To make full use of this local image appearance, we propose to solve a new task: human limb detection. We combine it with human joint localization in one deep convolutional neural network. After getting coarse results, we employ a graphical model to remove false positive detections. Besides, shallow and deep features are combined in this model. We evaluate our method on the FLIC and LSP datasets. The experiments results show the effectiveness of our method.

While feedforward deep convolutional neural networks (CNNs) have been a great success in computer vision, it is important to note that the human visual cortex generally contains more feedback than feedforward connections. In this paper, we will briefly introduce the background of feedbacks in the human visual cortex, which motivates us to develop a computational feedback mechanism in deep neural networks. In addition to the feedforward inference in traditional neural networks, a feedback loop is introduced to infer the activation status of hidden layer neurons according to the "goal" of the network, e.g., high-level semantic labels. We analogize this mechanism as "Look and Think Twice." The feedback networks help better visualize and understand how deep neural networks work, and capture visual attention on expected objects, even in images with cluttered background and multiple objects. Experiments on ImageNet dataset demonstrate its effectiveness in solving tasks such as image classification and object localization.

In this paper, we address the task of learning novel visual concepts, and their interactions with other concepts, from a few images with sentence descriptions. Using linguistic context and visual features, our method is able to efficiently hypothesize the semantic meaning of new words and add them to its word dictionary so that they can be used to describe images which contain these novel concepts. Our method has an image captioning module based on [38] with several improvements. In particular, we propose a transposed weight sharing scheme, which not only improves performance on image captioning, but also makes the model more suitable for the novel concept learning task. We propose methods to prevent overfitting the new concepts. In addition, three novel concept datasets are constructed for this new task, and are publicly available on the project page. In the experiments, we show that our method effectively learns novel visual concepts from a few examples without disturbing the previously learned concepts. The project page is: www.stat.ucla.edu/junhua. mao/projects/child_learning.html.

Existing methods on video-based action recognition are generally view-dependent, i.e., performing recognition from the same views seen in the training data. We present a novel multiview spatio-temporal and-or graph (MST-AOG) representation for cross-view action recognition, i.e., the recognition is performed on the video from an unknown and unseen view. As a compositional model, MST-AOG compactly represents the hierarchical combinatorial structures of cross-view actions by explicitly modeling the geometry, appearance and motion variations. This paper proposes effective methods to learn the structure and parameters of MST-AOG. The inference based on MST-AOG enables action recognition from novel views. The training of MST-AOG takes advantage of the 3D human skeleton data obtained from Kinect cameras to avoid annotating enormous multi-view video frames, which is error-prone and time-consuming, but the recognition does not need 3D information and is based on 2D video input. A new Multiview Action3D dataset has been created and will be released. Extensive experiments have demonstrated that this new action representation significantly improves the accuracy and robustness for cross-view action recognition on 2D videos.

Learning fine-grained image similarity is a challenging task. It needs to capture between-class and within-class image differences. This paper proposes a deep ranking model that employs deep learning techniques to learn similarity metric directly from images. It has higher learning capability than models based on hand-crafted features. A novel multiscale network structure has been developed to describe the images effectively. An efficient triplet sampling algorithm is also proposed to learn the model with distributed asynchronized stochastic gradient. Extensive experiments show that the proposed algorithm outperforms models based on hand-crafted visual features and deep classification models.

This paper presents a novel approach to cross-view action recognition. Traditional cross-view action recognition methods typically rely on local appearance/motion features. In this paper, we take advantage of the recent developments of depth cameras to build a more discriminative cross-view action representation. In this representation, an action is characterized by the spatio-temporal configuration of 3D Poselets, which are discriminatively discovered with a novel Poselet mining algorithm and can be detected with view-invariant 3D Poselet detectors. The Kinect skeleton is employed to facilitate the 3D Poselet mining and 3D Poselet detectors learning, but the recognition is solely based on 2D video input. Extensive experiments have demonstrated that this new action representation significantly improves the accuracy and robustness for cross-view action recognition.

Human action recognition is an important yet challenging task. Human actions usually involve human-object interactions, highly articulated motions, high intra-class variations, and complicated temporal structures. The recently developed commodity depth sensors open up new possibilities of dealing with this problem by providing 3D depth data of the scene. This information not only facilitates a rather powerful human motion capturing technique, but also makes it possible to efficiently model human-object interactions and intra-class variations. In this paper, we propose to characterize the human actions with a novel actionlet ensemble model, which represents the interaction of a subset of human joints. The proposed model is robust to noise, invariant to translational and temporal misalignment, and capable of characterizing both the human motion and the human-object interactions. We evaluate the proposed approach on three challenging action recognition datasets captured by Kinect devices, a multiview action recognition dataset captured with Kinect device, and a dataset captured by a motion capture system. The experimental evaluations show that the proposed approach achieves superior performance to the state-of-the-art algorithms.

Temporal misalignment and duration variation in video actions largely influence the performance of action recognition, but it is very difficult to specify effective temporal alignment on action sequences. To address this challenge, this paper proposes a novel discriminative learning-based temporal alignment method, called maximum margin temporal warping (MMTW), to align two action sequences and measure their matching score. Based on the latent structure SVM formulation, the proposed MMTW method is able to learn a phantom action template to represent an action class for maximum discrimination against other classes. The recognition of this action class is based on the associated learned alignment of the input action. Extensive experiments on five benchmark datasets have demonstrated that this MMTW model is able to significantly promote the accuracy and robustness of action recognition under temporal misalignment and variations.

Regularizing motion field is critical to achieve accurate estimation of the motion field. As the motion field may include discontinuity (e.g., at the motion boundaries), traditional smoothness regularization may not work well. Among many approaches to handling motion discontinuity, recent attempts pursued a sparse representation of the motion field for regularization, and achieved quite encouraging results. However, statistics show that these methods tend to over-sparsify the motion field, and thus confronted by the non-sparse noise in practice. In this paper, we propose to decompose the motion field into sparse and non-sparse components for the motion boundaries and small universal noises, respectively. This separation approach regularizes these two sources differently. We propose a novel and efficient optimization algorithm to solve this problem. In addition, our study reveals the in-depth connection between this noise separation approach and the influence function approach in robust statistics. We validate and evaluate our new approach on the Middlebury benchmark, and have achieved outstanding testing performance.

Human action recognition is an important yet challenging task. The recently developed commodity depth sensors open up new possibilities of dealing with this problem but also present some unique challenges. The depth maps captured by the depth cameras are very noisy and the 3D positions of the tracked joints may be completely wrong if serious occlusions occur, which increases the intra-class variations in the actions. In this paper, an actionlet ensemble model is learnt to represent each action and to capture the intra-class variance. In addition, novel features that are suitable for depth data are proposed. They are robust to noise, invariant to translational and temporal misalignments, and capable of characterizing both the human motion and the human-object interactions. The proposed approach is evaluated on two challenging action recognition datasets captured by commodity depth cameras, and another dataset captured by a MoCap system. The experimental evaluations show that the proposed approach achieves superior performance to the state of the art algorithms.

The popular bag of words approach for action recognition is based on the classifying quantized local features density. This approach focuses excessively on the local features but discards all information about the interactions among them. Local features themselves may not be discriminative enough, but combined with their contexts, they can be very useful for the recognition of some actions. In this paper, we present a novel representation that captures contextual interactions between interest points, based on the density of all features observed in each interest point's mutliscale spatio-temporal contextual domain. We demonstrate that augmenting local features with our contextual feature significantly improves the recognition performance.