CVPR 2017 论文列表

This paper addresses the problem of depth estimation from a single still image. Inspired by recent works on multi-scale convolutional neural networks (CNN), we propose a deep model which fuses complementary information derived from multiple CNN side outputs. Different from previous methods, the integration is obtained by means of continuous Conditional Random Fields (CRFs). In particular, we propose two different variations, one based on a cascade of multiple CRFs, the other on a unified graphical model. By designing a novel CNN implementation of mean-field updates for continuous CRFs, we show that both proposed models can be regarded as sequential deep networks and that training can be performed end-to-end. Through extensive experimental evaluation we demonstrate the effectiveness of the proposed approach and establish new state of the art results on publicly available datasets.

We propose Quadruplet Convolutional Neural Networks (Quad-CNN) for multi-object tracking, which learn to associate object detections across frames using quadruplet losses. The proposed networks consider target appearances together with their temporal adjacencies for data association. Unlike conventional ranking losses, the quadruplet loss enforces an additional constraint that makes temporally adjacent detections more closely located than the ones with large temporal gaps. We also employ a multi-task loss to jointly learn object association and bounding box regression for better localization. The whole network is trained end-to-end. For tracking, the target association is performed by minimax label propagation using the metric learned from the proposed network. We evaluate performance of our multi-object tracking algorithm on public MOT Challenge datasets, and achieve outstanding results.

Mean Field inference is central to statistical physics. It has attracted much interest in the Computer Vision community to efficiently solve problems expressible in terms of large Conditional Random Fields. However, since it models the posterior probability distribution as a product of marginal probabilities, it may fail to properly account for important dependencies between variables. We therefore replace the fully factorized distribution of Mean Field by a weighted mixture of such distributions, that similarly minimizes the KL-Divergence to the true posterior. By introducing two new ideas, namely, conditioning on groups of variables instead of single ones and using a parameter of the conditional random field potentials, that we identify to the temperature in the sense of statistical physics to select such groups, we can perform this minimization efficiently. Our extension of the clamping method proposed in previous works allows us to both produce a more descriptive approximation of the true posterior and, inspired by the diverse MAP paradigms, fit a mixture of Mean Field approximations. We demonstrate that this positively impacts real-world algorithms that initially relied on mean fields.

Inspired by the recent success of text-based question answering, visual question answering (VQA) is proposed to automatically answer natural language questions with the reference to a given image. Compared with text-based QA, VQA is more challenging because the reasoning process on visual domain needs both effective semantic embedding and fine-grained visual understanding. Existing approaches predominantly infer answers from the abstract low-level visual features, while neglecting the modeling of high-level image semantics and the rich spatial context of regions. To solve the challenges, we propose a multi-level attention network for visual question answering that can simultaneously reduce the semantic gap by semantic attention and benefit fine-grained spatial inference by visual attention. First, we generate semantic concepts from high-level semantics in convolutional neural networks (CNN) and select those question-related concepts as semantic attention. Second, we encode region-based middle-level outputs from CNN into spatially-embedded representation by a bidirectional recurrent neural network, and further pinpoint the answer-related regions by multiple layer perceptron as visual attention. Third, we jointly optimize semantic attention, visual attention and question embedding by a softmax classifier to infer the final answer. Extensive experiments show the proposed approach outperforms the-state-of-arts on two challenging VQA datasets.

In this paper, we propose to incorporate convolutional neural networks with a multi-context attention mechanism into an end-to-end framework for human pose estimation. We adopt stacked hourglass networks to generate attention maps from features at multiple resolutions with various semantics. The Conditional Random Field (CRF) is utilized to model the correlations among neighboring regions in the attention map. We further combine the holistic attention model, which focuses on the global consistency of the full human body, and the body part attention model, which focuses on detailed descriptions for different body parts. Hence our model has the ability to focus on different granularity from local salient regions to global semantic consistent spaces. Additionally, we design novel Hourglass Residual Units (HRUs) to increase the receptive field of the network. These units are extensions of residual units with a side branch incorporating filters with larger receptive field, hence features with various scales are learned and combined within the HRUs. The effectiveness of the proposed multi-context attention mechanism and the hourglass residual units is evaluated on two widely used human pose estimation benchmarks. Our approach outperforms all existing methods on both benchmarks over all the body parts. Code has been made publicly available.

One-shot learning is a challenging problem where the aim is to recognize a class identified by a single training image. Given the practical importance of one-shot learning, it seems surprising that the rich information present in the class tag itself has largely been ignored. Most existing approaches restrict the use of the class tag to finding similar classes and transferring classifiers or metrics learned thereon. We demonstrate here, in contrast, that the class tag can inform one-shot learning as a guide to visual attention on the training image for creating the image representation. This is motivated by the fact that human beings can better interpret a training image if the class tag of the image is understood. Specifically, we design a neural network architecture which takes the semantic embedding of the class tag to generate attention maps and uses those attention maps to create the image features for one-shot learning. Note that unlike other applications, our task requires that the learned attention generator can be generalized to novel classes. We show that this can be realized by representing class tags with distributed word embeddings and learning the attention map generator from an auxiliary training set. Also, we design a multiple-attention scheme to extract richer information from the exemplar image and this leads to substantial performance improvement. Through comprehensive experiments, we show that the proposed approach leads to superior performance over the baseline methods.

In a transfer-learning scheme, the intermediate layers of a pre-trained CNN are employed as universal image representation to tackle many visual classification problems. The current trend to generate such representation is to learn a CNN on a large set of images labeled among the most specific categories. Such processes ignore potential relations between categories, as well as the categorical-levels used by humans to classify. In this paper, we propose Multi Categorical-Level Networks (MuCaLe-Net) that include human-categorization knowledge into the CNN learning process. A MuCaLe-Net separates generic categories from each other while it independently distinguishes specific ones. It thereby generates different features in the intermediate layers that are complementary when combined together. Advantageously, our method does not require additive data nor annotation to train the network. The extensive experiments over four publicly available benchmarks of image classification exhibit state-of-the-art performances.

In this paper, we present a novel and general network structure towards accelerating the inference process of convolutional neural networks, which is more complicated in network structure yet with less inference complexity. The core idea is to equip each original convolutional layer with another low-cost collaborative layer (LCCL), and the element-wise multiplication of the ReLU outputs of these two parallel layers produces the layer-wise output. The combined layer is potentially more discriminative than the original convolutional layer, and its inference is faster for two reasons: 1) the zero cells of the LCCL feature maps will remain zero after element-wise multiplication, and thus it is safe to skip the calculation of the corresponding high-cost convolution in the original convolutional layer; 2) LCCL is very fast if it is implemented as a 1*1 convolution or only a single filter shared by all channels. Extensive experiments on the CIFAR-10, CIFAR-100 and ILSCRC-2012 benchmarks show that our proposed network structure can accelerate the inference process by 32% on average with negligible performance drop.

Recently, skeleton based action recognition gains more popularity due to cost-effective depth sensors coupled with real-time skeleton estimation algorithms. Traditional approaches based on handcrafted features are limited to represent the complexity of motion patterns. Recent methods that use Recurrent Neural Networks (RNN) to handle raw skeletons only focus on the contextual dependency in the temporal domain and neglect the spatial configurations of articulated skeletons. In this paper, we propose a novel two-stream RNN architecture to model both temporal dynamics and spatial configurations for skeleton based action recognition. We explore two different structures for the temporal stream: stacked RNN and hierarchical RNN. Hierarchical RNN is designed according to human body kinematics. We also propose two effective methods to model the spatial structure by converting the spatial graph into a sequence of joints. To improve generalization of our model, we further exploit 3D transformation based data augmentation techniques including rotation and scaling transformation to transform the 3D coordinates of skeletons during training. Experiments on 3D action recognition benchmark datasets show that our method brings a considerable improvement for a variety of actions, i.e., generic actions, interaction activities and gestures.

First-person videos have unique characteristics such as heavy egocentric motion, strong preceding events, salient transitional activities and post-event impacts. Action recognition methods designed for third person videos may not optimally represent actions captured by first-person videos. We propose a method to represent the high level dynamics of sub-events in first-person videos by dynamically pooling features of sub-intervals of time series using a temporal feature pooling function. The sub-event dynamics are then temporally aligned to make a new series. To keep track of how the sub-event dynamics evolve over time, we recursively employ the Fast Fourier Transform on a pyramidal temporal structure. The Fourier coefficients of the segment define the overall video representation. We perform experiments on two existing benchmark first-person video datasets which have been captured in a controlled environment. Addressing this gap, we introduce a new dataset collected from YouTube which has a larger number of classes and a greater diversity of capture conditions thereby more closely depicting real-world challenges in first-person video analysis. We compare our method to state-of-the-art first person and generic video recognition algorithms. Our method consistently outperforms the nearest competitors by 10.3%, 3.3% and 11.7% respectively on the three datasets.

People often refer to entities in an image in terms of their relationships with other entities. For example, "the black cat sitting under the table" refers to both a "black cat" entity and its relationship with another "table" entity. Understanding these relationships is essential for interpreting and grounding such natural language expressions. Most prior work focuses on either grounding entire referential expressions holistically to one region, or localizing relationships based on a fixed set of categories. In this paper we instead present a modular deep architecture capable of analyzing referential expressions into their component parts, identifying entities and relationships mentioned in the input expression and grounding them all in the scene. We call this approach Compositional Modular Networks (CMNs): a novel architecture that learns linguistic analysis and visual inference end-to-end. Our approach is built around two types of neural modules that inspect local regions and pairwise interactions between regions. We evaluate CMNs on multiple referential expression datasets, outperforming state-of-the-art approaches on all tasks.

The inherent noise in the observed (e.g., scanned) binary document image degrades the image quality and harms the compression ratio through breaking the pattern repentance and adding entropy to the document images. In this paper, we design a cost function in Bayesian framework with dictionary learning. Minimizing our cost function produces a restored image which has better quality than that of the observed noisy image, and a dictionary for representing and encoding the image. After the restoration, we use this dictionary (from the same cost function) to encode the restored image following the symbol-dictionary framework by JBIG2 standard with the lossless mode. Experimental results with a variety of document images demonstrate that our method improves the image quality compared with the observed image, and simultaneously improves the compression ratio. For the test images with synthetic noise, our method reduces the number of flipped pixels by 48.2% and improves the compression ratio by 36.36% as compared with the best encoding methods. For the test images with real noise, our method visually improves the image quality, and outperforms the cutting-edge method by 28.27% in terms of the compression ratio.

Affordable sensors lead to an increasing interest in acquiring and modeling data with multiple modalities. Learning from multiple modalities has shown to significantly improve performance in object recognition. However, in practice it is common that the sensing equipment experiences unforeseeable malfunction or configuration issues, leading to corrupted data with missing modalities. Most existing multi-modal learning algorithms could not handle missing modalities, and would discard either all modalities with missing values or all corrupted data. To leverage the valuable information in the corrupted data, we propose to impute the missing data by leveraging the relatedness among different modalities. Specifically, we propose a novel Cascaded Residual Autoencoder (CRA) to impute missing modalities. By stacking residual autoencoders, CRA grows iteratively to model the residual between the current prediction and original data. Extensive experiments demonstrate the superior performance of CRA on both the data imputation and the object recognition task on imputed data.

Given a convolutional neural network (CNN) that is pre-trained for object classification, this paper proposes to use active question-answering to semanticize neural patterns in conv-layers of the CNN and mine part concepts. For each part concept, we mine neural patterns in the pre-trained CNN, which are related to the target part, and use these patterns to construct an And-Or graph (AOG) to represent a four-layer semantic hierarchy of the part. As an interpretable model, the AOG associates different CNN units with different explicit object parts. We use an active human-computer communication to incrementally grow such an AOG on the pre-trained CNN as follows. We allow the computer to actively identify objects, whose neural patterns cannot be explained by the current AOG. Then, the computer asks human about the unexplained objects, and uses the answers to automatically discover certain CNN patterns corresponding to the missing knowledge. We incrementally grow the AOG to encode new knowledge discovered during the active-learning process. In experiments, our method exhibits high learning efficiency. Our method uses about 1/6--1/3 of the part annotations for training, but achieves similar or better part-localization performance than fast-RCNN methods.

We present a general framework and method for detection of an object in a video based on apparent motion. The object moves relative to background motion at some unknown time in the video, and the goal is to detect and segment the object as soon it moves in an online manner. Due to unreliability of motion between frames, more than two frames are needed to reliably detect the object. Our method is designed to detect the object(s) with minimum delay, i.e., frames after the object moves, constraining the false alarms. Experiments on a new extensive dataset for moving object detection show that our method achieves less delay for all false alarm constraints than existing state-of-the-art.

In domain adaptation, maximum mean discrepancy (MMD) has been widely adopted as a discrepancy metric between the distributions of source and target domains. However, existing MMD-based domain adaptation methods generally ignore the changes of class prior distributions, i.e., class weight bias across domains. This remains an open problem but ubiquitous for domain adaptation, which can be caused by changes in sample selection criteria and application scenarios. We show that MMD cannot account for class weight bias and results in degraded domain adaptation performance. To address this issue, a weighted MMD model is proposed in this paper. Specifically, we introduce class-specific auxiliary weights into the original MMD for exploiting the class prior probability on source and target domains, whose challenge lies in the fact that the class label in target domain is unavailable. To account for it, our proposed weighted MMD model is defined by introducing an auxiliary weight for each class in the source domain, and a classification EM algorithm is suggested by alternating between assigning the pseudo-labels, estimating auxiliary weights and updating model parameters. Extensive experiments demonstrate the superiority of our weighted MMD over conventional MMD for domain adaptation.

Current CNN based object detectors need initialization from pre-trained ImageNet classification models, which are usually time-consuming. In this paper, we present a fully convolutional feature mimic framework to train very efficient CNN based detectors, which do not need ImageNet pre-training and achieve competitive performance as the large and slow models. We add supervision from high-level features of the large networks in training to help the small network better learn object representation. More specifically, we conduct a mimic method for the features sampled from the entire feature map and use a transform layer to map features from the small network onto the same dimension of the large network. In training the small network, we optimize the similarity between features sampled from the same region on the feature maps of both networks. Extensive experiments are conducted on pedestrian and common object detection tasks using VGG, Inception and ResNet. On both Caltech and Pascal VOC, we show that the modified 2.5x accelerated Inception network achieves competitive performance as the full Inception Network. Our faster model runs at 80 FPS for a 1000x1500 large input with only a minor degradation of performance on Caltech.

We introduce a new fashion search protocol where attribute manipulation is allowed within the interaction between users and search engines, e.g. manipulating the color attribute of the clothing from red to blue. It is particularly useful for image-based search when the query image cannot perfectly match user's expectation of the desired product. To build such a search engine, we propose a novel memory-augmented Attribute Manipulation Network (AMNet) which can manipulate image representation at the attribute level. Given a query image and some attributes that need to modify, AMNet can manipulate the intermediate representation encoding the unwanted attributes and change them to the desired ones through following four novel components: (1) a dual-path CNN architecture for discriminative deep attribute representation learning; (2) a memory block with an internal memory and a neural controller for prototype attribute representation learning and hosting; (3) an attribute manipulation network to modify the representation of the query image with the prototype feature retrieved from the memory block; (4) a loss layer which jointly optimizes the attribute classification loss and a triplet ranking loss over triplet images for facilitating precise attribute manipulation and image retrieving. Extensive experiments conducted on two large-scale fashion search datasets, i.e. DARN and DeepFashion, have demonstrated that AMNet is able to achieve remarkably good performance compared with well-designed baselines in terms of effectiveness of attribute manipulation and search accuracy.

Thin structures such as fence, grass and vessels are common in photography and scientific imaging. They exhibit complex 3D structures with sharp depth variations/discontinuities and mutual occlusions. In this paper, we develop a method to estimate the occlusion matte and depths of thin structures from a focal image stack, which is obtained either by varying the focus/aperture of the lens or computed from a one-shot light field image. We propose an image formation model that explicitly describes the spatially varying optical blur and mutual occlusions for structures located at different depths. Based on the model, we derive an efficient MCMC inference algorithm that enables direct and analytical computations of the iterative update for the model/images without re-rendering images in the sampling process. Then, the depths of the thin structures are recovered using gradient descent with the differential terms computed using the image formation model. We apply the proposed method to scenes at both macro and micro scales. For macro-scale, we evaluate our method on scenes with complex 3D thin structures such as tree branches and grass. For micro-scale, we apply our method to in-vivo microscopic images of micro-vessels with diameters less than 50 um. To our knowledge, the proposed method is the first approach to reconstruct the 3D structures of micro-vessels from non-invasive in-vivo image measurements.