ECCV 2018 论文列表

In this paper, we make two contributions to unsupervised domain adaptation (UDA) using the convolutional neural network (CNN). First, our approach transfers knowledge in all the convolutional layers through attention alignment. Most previous methods align high-level representations, e.g., activations of the fully connected (FC) layer. In these methods, however, the convolutional layers which underpin critical low-level domain knowledge cannot be updated directly towards reducing domain discrepancy. Specifically, we assume that the discriminative regions in an image are relatively invariant to image style changes. Based on this assumption, we propose an attention alignment scheme on all the target convolutional layers to uncover the knowledge shared by the source domain. Second, we estimate the posterior label distribution of the unlabeled data for target network training. Previous methods, which iteratively update the pseudo labels by the target network and refine the target network by the updated pseudo labels, are vulnerable to label estimation errors. Instead, our approach uses category distribution to calculate the cross-entropy loss for training, thereby ameliorating the error accumulation of the estimated labels. The two contributions allow our approach to outperform the state-of-the-art methods by +2.6% on the Office-31 dataset.

Facial action unit (AU) detection and face alignment are two highly correlated tasks since facial landmarks can provide precise AU locations to facilitate the extraction of meaningful local features for AU detection. Most existing AU detection works often treat face alignment as a preprocessing and handle the two tasks independently. In this paper, we propose a novel end-to-end deep learning framework for joint AU detection and face alignment, which has not been explored before. In particular, multi-scale shared features are learned firstly, and high-level features of face alignment are fed into AU detection. Moreover, to extract precise local features, we propose an adaptive attention learning module to refine the attention map of each AU adaptively. Finally, the assembled local features are integrated with face alignment features and global features for AU detection. Experiments on BP4D and DISFA benchmarks demonstrate that our framework significantly outperforms the state-of-the-art methods for AU detection.

Many different deep networks have been used to approximate, accelerate or improve traditional image operators, such as image smoothing, super-resolution and denoising. Among these traditional operators, many contain parameters which need to be tweaked to obtain the satisfactory results, which we refer to as "parameterized image operators''. However, most existing deep networks trained for these operators are only designed for one specific parameter configuration, which does not meet the needs of real scenarios that usually require flexible parameters settings. To overcome this limitation, we propose a new decouple learning algorithm to learn from the operator parameters to dynamically adjust the weights of a deep network for image operators, denoted as the base network. The learned algorithm is formed as another network, namely the weight learning network, which can be end-to-end jointly trained with the base network. Experiments demonstrate that the proposed framework can be successfully applied to many traditional parameterized image operators. We provide more analysis to better understand the proposed framework, which may inspire more promising research in this direction. Our codes and models have been released in https://github.com/fqnchina/DecoupleLearning.

Blind image deblurring is a challenging problem due to its ill-posed nature, of which the success is closely related to a proper image prior. Although a large number of sparsity-based priors, such as the sparse gradient prior, have been successfully applied for blind image deblurring, they inherently suffer from several drawbacks, limiting their applications. Existing sparsity-based priors are usually rooted in modeling the response of images to some specific filters (e.g., image gradients), which are insufficient to capture the complicated image structures. Moreover, the traditional sparse priors or regularizations model the filter response (e.g., image gradients) independently and thus fail to depict the long range correlation among them. To address the above issues, we present a novel image prior for image deblurring based on a Super-Gaussian field model with adaptive structures. Instead of modeling the response of the fixed short-term filters, the proposed Super-Gaussian fields capture the complicated structures in natural images by integrating potentials on all cliques (e.g., centring at each pixel) into a joint probabilistic distribution. Considering that the fixed filters in different scales are impractical for the coarse-to-fine framework of image deblurring, we define each potential function as a super-Gaussian distribution. Through this definition, the partition function, the curse for traditional MRFs, can be theoretically ignored, and all model parameters of the proposed Super-Gaussian fields can be data-adaptively learned and inferred from the blurred observation with a variational framework. Extensive experiments on both blind deblurring and non-blind deblurring demonstrate the effectiveness of the proposed method.

Deep convolutional neural networks have liberated its extraordinary power on various tasks. However, it is still very challenging to deploy state-of-the-art models into real-world applications due to their high computational complexity. How can we design a compact and effective network without massive experiments and expert knowledge? In this paper, we propose a simple and effective framework to learn and prune deep models in an end-to-end manner. In our framework, a new type of parameter -- scaling factor is first introduced to scale the outputs of specific structures, such as neurons, groups or residual blocks. Then we add sparsity regularizations on these factors, and solve this optimization problem by a modified stochastic Accelerated Proximal Gradient (APG) method. By forcing some of the factors to zero, we can safely remove the corresponding structures, thus prune the unimportant parts of a CNN. Comparing with other structure selection methods that may need thousands of trials or iterative fine-tuning, our method is trained fully end-to-end in one training pass without bells and whistles. We evaluate our method, Sparse Structure Selection with several state-of-the-art CNNs, and demonstrate very promising results with adaptive depth and width selection. Code is available at: https://github.com/huangzehao/sparse-structure-selection.

The ability to anticipate the future is essential when making real time critical decisions, provides valuable information to understand dynamic natural scenes, and can help unsupervised video representation learning. State-of-art video prediction is based on complex architectures that need to learn large numbers of parameters, are potentially hard to train, slow to run, and may produce blurry predictions. In this paper, we introduce DYAN, a novel network with very few parameters and easy to train, which produces accurate, high quality frame predictions, significantly faster than previous approaches. DYAN owes its good qualities to its encoder and decoder, which are designed following concepts from systems identification theory and exploit the dynamics-based invariants of the data. Extensive experiments using several standard video datasets show that DYAN is superior generating frames and that it generalizes well across domains.

Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performances in applications such as image classification and language modeling. However, these techniques typically ignore device-related objectives such as inference time, memory usage, and power consumption. Optimizing neural architecture for device-related objectives is immensely crucial for deploying deep networks on portable devices with limited computing resources. We propose DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures, optimizing for both device-related (e.g., inference time and memory usage) and device-agnostic (e.g., accuracy and model size) objectives. DPP-Net employs a compact search space inspired by current state-of-the-art mobile CNNs, and further improves search efficiency by adopting progressive search (Liu et al. 2017). Experimental results on CIFAR-10 are poised to demonstrate the effectiveness of Pareto-optimal networks found by DPP-Net, for three different devices: (1) a workstation with Titan X GPU, (2) NVIDIA Jetson TX1 embedded system, and (3) mobile phone with ARM Cortex-A53. Compared to CondenseNet and NASNet (Mobile), DPP-Net achieves better performances: higher accuracy and shorter inference time on various devices. Additional experimental results show that models found by DPP-Net also achieve considerably-good performance on ImageNet as well.

We present a scalable approach for Detecting Objects by transferring Common-sense Knowledge (DOCK) from source to target categories. In our setting, the training data for the source categories have bounding box annotations, while those for the target categories only have image-level annotations. Current state-of-the-art approaches focus on image-level visual or semantic similarity to adapt a detector trained on the source categories to the new target categories. In contrast, our key idea is to (i) use similarity not at the image-level, but rather at the region-level, and (ii) leverage richer common-sense (based on attribute, spatial, etc.) to guide the algorithm towards learning the correct detections. We acquire such common-sense cues automatically from readily-available knowledge bases without any extra human effort. On the challenging MS COCO dataset, we find that common-sense knowledge can substantially improve detection performance over existing transfer-learning baselines.

We propose a novel discrete Fourier transform-based pooling layer for convolutional neural networks. The DFT magnitude pooling replaces the traditional max/average pooling layer between the convolution and fully-connected layers to retain translation invariance and shape preserving (aware of shape difference) properties based on the shift theorem of the Fourier transform. Thanks to the ability to handle image misalignment while keeping important structural information in the pooling stage, the DFT magnitude pooling improves the classification accuracy significantly. In addition, we propose the DFT+ method for ensemble networks using the middle convolution layer outputs. The proposed methods are extensively evaluated on various classification tasks using the ImageNet, CUB 2010-2011, MIT Indoors, Caltech 101, FMD and DTD datasets. The AlexNet, VGG-VD 16, Inception-v3, and ResNet are used as the base networks, upon which DFT and DFT+ methods are implemented. Experimental results show that the proposed methods improve the classification performance in all networks and datasets.

We present an unsupervised learning framework for simultaneously training single-view depth prediction and optical flow estimation models using unlabeled video sequences. Existing unsupervised methods often exploit brightness constancy and spatial smoothness priors to train depth or flow models. In this paper, we propose to leverage geometric consistency as additional supervisory signals. Our core idea is that for rigid regions we can use the predicted scene depth and camera motion to synthesize 2D optical flow by backprojecting the induced 3D scene flow. The discrepancy between the rigid flow (from depth prediction and camera motion) and the estimated flow (from optical flow model) allows us to impose a cross-task consistency loss. While all the networks are jointly optimized during training, they can be applied independently at test time. Extensive experiments demonstrate that our depth and flow models compare favorably with state-of-the-art unsupervised methods.

Consumer depth sensors are more and more popular and come to our daily lives marked by its recent integration in the latest Iphone X. However, they still suffer from heavy noises which limit their applications. Although plenty of progresses have been made to reduce the noises and boost geometric details, due to the inherent illness and the real-time requirement, the problem is still far from been solved. We propose a cascaded Depth Denoising and Refinement Network (DDRNet) to tackle this problem by leveraging the multi-frame fused geometry and the accompanying high quality color image through a joint training strategy. The rendering equation is exploited in our network in an unsupervised manner. In detail, we impose an unsupervised loss based on the light transport to extract the high-frequency geometry. Experimental results indicate that our network achieves real-time single depth enhancement on various categories of scenes. Thanks to the well decoupling of the low and high frequency information in the cascaded network, we achieve superior performance over the state-of-the-art techniques.

Harvesting dense pixel-level annotations to train deep neural networks for semantic segmentation is extremely expensive and unwieldy at scale. While learning from synthetic data where labels are readily available sounds promising, performance degrades significantly when testing on novel realistic data due to domain discrepancies. We present Dual Channel-wise Alignment Networks (DCAN), a simple yet effective approach to reduce domain shift at both pixel-level and feature-level. Exploring statistics in each channel of CNN feature maps, our framework performs channel-wise feature alignment, which preserves spatial structures and semantic information, in both an image generator and a segmentation network. In particular, given an image from the source domain and unlabeled samples from the target domain, the generator synthesizes new images on-the-fly to resemble samples from the target domain in appearance and the segmentation network further refines high-level features before predicting semantic maps, both of which leverage feature statistics of sampled images from the target domain. Unlike much recent and concurrent work relying on adversarial training, our framework is lightweight and easy to train. Extensive experiments on adapting models trained on synthetic segmentation benchmarks to real urban scenes demonstrate the effectiveness of the proposed framework.

We present a simple yet efficient approach capable of training deep neural networks on large-scale weakly-supervised web images, which are crawled rawly from the Internet by using text queries, without any human annotation. We develop a principled learning strategy by leveraging curriculum learning, with the goal of handling massive amount of noisy labels and data imbalance effectively. We design a new learning curriculum by measuring the complexity of data using its distribution density in a feature space, and rank the complexity in an unsupervised manner. This allows for an efficient implementation of curriculum learning on large-scale web images, resulting in a high-performance CNN model, where the negative impact of noisy labels is reduced substantially. Importantly, we show by experiments that those images with highly noisy labels can surprisingly improve the generalization capability of model, by serving as a manner of regularization. Our approaches obtain the state-of-the-art performance on four benchmarks, including Webvision, ImageNet, Clothing-1M and Food-101. With an ensemble of multiple models, we achieve a top-5 error rate of 5.2% on the Webvision challenge cite{li2017webvision} for 1000-category classification, which is the top performance that surpasses other results by a large margin of about 50% relative error rate. Codes and models are available at: https://github.com/guoshengcv/CurriculumNet.

3D Convolutional Neural Networks are sensitive to transformations applied to their input. This is a problem because a voxelized version of a 3D object, and its rotated clone, will look unrelated to each other after passing through to the last layer of a network. Instead, an idealized model would preserve a meaningful representation of the voxelized object, while explaining the pose-difference between the two inputs. An equivariant representation vector has two components: the invariant identity part, and a discernable encoding of the transformation. Models that can't explain pose-differences risk ``diluting'' the representation, in pursuit of optimizing a classification or regression loss function. We introduce a Group Convolutional Neural Network with linear equivariance to translations and right angle rotations in three dimensions. We call this network CubeNet, reflecting its cube-like symmetry. By construction, this network helps preserve a 3D shape's global and local signature, as it is transformed through successive layers. We apply this network to a variety of 3D inference problems, achieving state-of-the-art on the ModelNet10 classification challenge, and comparable performance on the ISBI 2012 Connectome Segmentation Benchmark. To the best of our knowledge, this is the first 3D rotation equivariant CNN for voxel representations.

The Reference-based Super-resolution (RefSR) super-resolves a low-resolution (LR) image given an external high-resolution (HR) reference image, where the reference image and LR image share similar viewpoint but with significant resolution gap x8. Existing RefSR methods work in a cascaded way such as patch matching followed by synthesis pipeline with two independently defined objective functions, leading to the inter-patch misalignment, grid effect and inefficient optimization. To resolve these issues, we present CrossNet, an end-to-end and fully-convolutional deep neural network using cross-scale warping. Our network contains image encoders, cross-scale warping layers, and fusion decoder: the encoder serves to extract multi-scale features from both the LR and the reference images; the cross-scale warping layers spatially aligns the reference feature map with the LR feature map; the decoder finally aggregates feature maps from both domains to synthesize the HR output. Using cross-scale warping, our network is able to perform spatial alignment at pixel-level in an end-to-end fashion, which improves the existing schemes both in precision (around 2dB-4dB) and efficiency (more than 100 times faster).

Visual data and text data are composed of information at multiple granularities. A video can describe a complex scene that is composed of multiple clips or shots, where each depicts a semantically coherent event or action. Similarly, a paragraph may contain sentences with different topics, which collectively conveys a coherent message or story. In this paper, we investigate the modeling techniques for such hierarchical sequential data where there are correspondences across multiple modalities. Specifically, we introduce hierarchical sequence embedding (HSE), a generic model for embedding sequential data of different modalities into hierarchically semantic spaces, with either explicit or implicit correspondence information. We perform empirical studies on large-scale video and paragraph retrieval datasets and demonstrated superior performance by the proposed methods. Furthermore, we examine the effectiveness of our learned embeddings when applied to downstream tasks. We show its utility in zero-shot action recognition and video captioning.

Due to the complementary benefits of visible (RGB) and thermal infrared (T) data, RGB-T object tracking attracts more and more attention recently for boosting the performance under adverse illumination conditions. Existing RGB-T tracking methods usually localize a target object with a bounding box, in which the trackers or detectors is often affected by the inclusion of background clutter. To address this problem, this paper presents a novel approach to suppress background effects for RGB-T tracking. Our approach relies on a novel cross-modal manifold ranking algorithm. First, we integrate the soft cross-modality consistency into the ranking model which allows the sparse inconsistency to account for the different properties between these two modalities. Second, we propose an optimal query learning method to handle label noises of queries. In particular, we introduce an intermediate variable to represent the optimal labels, and formulate it as a $l_1$-optimization based sparse learning problem. Moreover, we propose a single unified optimization algorithm to solve the proposed model with stable and efficient convergence behavior. Finally, the ranking results are incorporated into the patch-based object features to address the background effects, and the structured SVM is then adopted to perform RGB-T tracking. Extensive experiments suggest that the proposed approach performs well against the state-of-the-art methods on large-scale benchmark datasets.

Cross-modal hashing enables similarity retrieval across different content modalities, such as searching relevant images in response to text queries. It provides with the advantages of computation efficiency and retrieval quality for multimedia retrieval. Hamming space retrieval enables efficient constant-time search that returns data items within a given Hamming radius to each query, by hash lookups instead of linear scan. However, Hamming space retrieval is ineffective in existing cross-modal hashing methods, subject to their weak capability of concentrating the relevant items to be within a small Hamming ball, while worse still, the Hamming distances between hash codes from different modalities are inevitably large due to the large heterogeneity across different modalities. This work presents Cross-Modal Hamming Hashing (CMHH), a novel deep cross-modal hashing approach that generates compact and highly concentrated hash codes to enable efficient and effective Hamming space retrieval. The main idea is to penalize significantly on similar cross-modal pairs with Hamming distance larger than the Hamming radius threshold, by designing a pairwise focal loss based on the exponential distribution. Extensive experiments demonstrate that CMHH can generate highly concentrated hash codes and achieve state-of-the-art cross-modal retrieval performance for both hash lookups and linear scan scenarios on three benchmark datasets, NUS-WIDE, MIRFlickr-25K, and IAPR TC-12.

Triplet loss, popular for metric learning, has made a great success in many computer vision tasks, such as fine-grained image classification, image retrieval, and face recognition. Considering that the number of triplets grows cubically with the size of training data, triplet mining is thus indispensable for efficiently training with triplet loss. However, in practice, the training is usually very sensitive to the selected triplets, e.g., it almost does not converge with randomly selected triplets and selecting hardest triplets also leads to bad local minima. We argue that the bias in sampling of triplets degrades the performance of learning with triplet loss. In this paper, we propose a new variant of triplet loss, which tries to reduce the bias in triplet sampling by adaptively correcting the distribution shift on sampled triplets. We refer to this new triplet loss as adapted triplet loss. We conduct a number of experiments on MNIST and Fashion-MNIST for image classification, and on CARS196, CUB200-2011, and Stanford Online Products for image retrieval. The experimental results demonstrate the effectiveness of the proposed method.