CVPR 2017 Papers

We present DeepNav, a Convolutional Neural Network (CNN) based algorithm for navigating large cities using locally visible street-view images. The DeepNav agent learns to reach its destination quickly by making the correct navigation decisions at intersections. We collect a large-scale dataset of street-view images organized in a graph where nodes are connected by roads. This dataset contains 10 city graphs and a total of more than 1 million street-view images. We propose 3 supervised learning approaches for the navigation task, and show how A* search in the city graph can be used to generate labels for the images. Our annotation process is fully automated using publicly available mapping services, and requires no human input. We evaluate the proposed DeepNav models on 4 held-out cities for navigating to 5 different types of destinations and show that our algorithms outperform previous work that uses hand-crafted features and Support Vector Regression (SVR).

Most contemporary approaches to instance segmentation use complex pipelines involving conditional random fields, recurrent neural networks, object proposals, or template matching schemes. In this paper, we present a simple yet powerful end-to-end convolutional neural network to tackle this task. Our approach combines intuitions from the classical watershed transform and modern deep learning to produce an energy map of the image where object instances are unambiguously represented as energy basins. We then perform a cut at a single energy level to directly yield connected components corresponding to object instances. Our model achieves more than double the performance over the state-of-the-art on the challenging Cityscapes Instance Level Segmentation task.

Compact coding has been widely applied to approximate nearest neighbor search for large-scale image retrieval, due to its computation efficiency and retrieval quality. This paper presents a compact coding solution with a focus on the deep learning to quantization approach, which improves retrieval quality by end-to-end representation learning and compact encoding and has already shown the superior performance over the hashing solutions for similarity retrieval. We propose Deep Visual-Semantic Quantization (DVSQ), which is the first approach to learning deep quantization models from labeled image data as well as the semantic information underlying general text domains. The main contribution lies in jointly learning deep visual-semantic embeddings and visual-semantic quantizers using carefully-designed hybrid networks and well-specified loss functions. DVSQ enables efficient and effective image retrieval by supporting maximum inner-product search, which is computed based on learned codebooks with fast distance table lookup. Comprehensive empirical evidence shows that DVSQ can generate compact binary codes and yield state-of-the-art similarity retrieval performance on standard benchmarks.

Recently, convolutional neural networks (CNN) have been successfully applied to view synthesis problems. However, such CNN-based methods can suffer from lack of texture details, shape distortions, or high computational complexity. In this paper, we propose a novel CNN architecture for view synthesis called "Deep View Morphing" that does not suffer from these issues. To synthesize a middle view of two input images, a rectification network first rectifies the two input images. An encoder-decoder network then generates dense correspondences between the rectified images and blending masks to predict the visibility of pixels of the rectified images in the middle view. A view morphing network finally synthesizes the middle view using the dense correspondences and blending masks. We experimentally show the proposed method significantly outperforms the state-of-the-art CNN-based view synthesis method.

Motion blur from camera shake is a major problem in videos captured by hand-held devices. Unlike single-image deblurring, video-based approaches can take advantage of the abundant information that exists across neighboring frames. As a result the best performing methods rely on the alignment of nearby frames. However, aligning images is a computationally expensive and fragile procedure, and methods that aggregate information must therefore be able to identify which regions have been accurately aligned and which have not, a task that requires high level scene understanding. In this work, we introduce a deep learning solution to video deblurring, where a CNN is trained end-to-end to learn how to accumulate information across frames. To train this network, we collected a dataset of real videos recorded with a high frame rate camera, which we use to generate synthetic motion blur for supervision. We show that the features learned from this dataset extend to deblurring motion blur that arises due to camera shake in a wide range of videos, and compare the quality of results to a number of other baselines.

Despite progress in visual perception tasks such as image classification and detection, computers still struggle to understand the interdependency of objects in the scene as a whole, e.g., relations between objects or their attributes. Existing methods often ignore global context cues capturing the interactions among different object instances, and can only recognize a handful of types by exhaustively training individual detectors for all possible relationships. To capture such global interdependency, we propose a deep Variation-structured Reinforcement Learning (VRL) framework to sequentially discover object relationships and attributes in the whole image. First, a directed semantic action graph is built using language priors to provide a rich and compact representation of semantic correlations between object categories, predicates, and attributes. Next, we use a variation-structured traversal over the action graph to construct a small, adaptive action set for each step based on the current state and historical actions. In particular, an ambiguity-aware object mining scheme is used to resolve semantic ambiguity among object categories that the object detector fails to distinguish. We then make sequential predictions using a deep RL framework, incorporating global context cues and semantic embeddings of previously extracted phrases in the state vector. Our experiments on the Visual Relationship Detection (VRD) dataset and the large-scale Visual Genome dataset validate the superiority of VRL, which can achieve significantly better detection results on datasets involving thousands of relationship and attribute types. We also demonstrate that VRL is able to predict unseen types embedded in our action graph by learning correlations on shared graph nodes.

Unsupervised learning of visual similarities is of paramount importance to computer vision, particularly due to lacking training data for fine-grained similarities. Deep learning of similarities is often based on relationships between pairs or triplets of samples. Many of these relations are unreliable and mutually contradicting, implying inconsistencies when trained without supervision information that relates different tuples or triplets to each other. To overcome this problem, we use local estimates of reliable (dis-)similarities to initially group samples into compact surrogate classes and use local partial orders of samples to classes to link classes to each other. Similarity learning is then formulated as a partial ordering task with soft correspondences of all samples to classes. Adopting a strategy of self-supervision, a CNN is trained to optimally represent samples in a mutually consistent manner while updating the classes. The similarity learning and grouping procedure are integrated in a single model and optimized jointly. The proposed unsupervised approach shows competitive performance on detailed pose estimation and object classification.

The CNN-encoding of features from entire videos for the representation of human actions has rarely been addressed. Instead, CNN work has focused on approaches to fuse spatial and temporal networks, but these were typically limited to processing shorter sequences. We present a new video representation, called temporal linear encoding (TLE) and embedded inside of CNNs as a new layer, which captures the appearance and motion throughout entire videos. It encodes this aggregated information into a robust video feature representation, via end-to-end learning. Advantages of TLEs are: (a) they encode the entire video into a compact feature representation, learning the semantics and a discriminative feature space; (b) they are applicable to all kinds of networks like 2D and 3D CNNs for video classification; and (c) they model feature interactions in a more expressive way and without loss of information. We conduct experiments on two challenging human action datasets: HMDB51 and UCF101. The experiments show that TLE outperforms current state-of-the-art methods on both datasets.

We propose a Deep Texture Encoding Network (TEN) with a novel Encoding Layer integrated on top of convolutional layers, which ports the entire dictionary learning and encoding pipeline into a single model. Current methods build from distinct components, using standard encoders with separate off-the-shelf features such as such as SIFT descriptors or pre-trained CNN features for material recognition. Our new approach provides an end-to-end learning framework, where the inherent visual vocabularies are learned directly from the loss function. That is, the features, dictionaries and the encoding representation for the classifier are all learned simultaneously. The representation is orderless and therefore is particularly useful for material and texture recognition. This Encoding Layer generalizes robust residual encoders such as VLAD and Fisher Vectors, and has the property of discarding domain specific information which makes the learned convolutional features easier to transfer. Additionally, joint training using multiple datasets of varied sizes and class labels is supported resulting in increased recognition performance. The experimental results show superior performance as compared to state-of-the-art methods using gold-standard databases such as MINC-2500, Flicker Material Database, KTH-TIPS-2b, and a new ground terrain multiview database. The source code for the complete system will be publicly available upon publication.

Monocular 3D object parsing is highly desirable in various scenarios including occlusion reasoning and holistic scene interpretation. We present a deep convolutional neural network (CNN) architecture to localize semantic parts in 2D image and 3D space while inferring their visibility states, given a single RGB image. Our key insight is to exploit domain knowledge to regularize the network by deeply supervising its hidden layers, in order to sequentially infer intermediate concepts associated with the final task. To acquire training data in desired quantities with ground truth 3D shape and relevant concepts, we render 3D object CAD models to generate large-scale synthetic data and simulate challenging occlusion configurations between objects. We train the network only on synthetic data and demonstrate state-of-the-art performances on real image benchmarks including an extended version of KITTI, PASCAL VOC, PASCAL3D+ and IKEA for 2D and 3D keypoint localization and instance segmentation. The empirical results substantiate the utility of our deep supervision scheme by demonstrating effective transfer of knowledge from synthetic data to real images, resulting in less overfitting compared to standard end-to-end training.

We consider the task of automated estimation of facial expression intensity. This involves estimation of multiple output variables (facial action units --- AUs) that are structurally dependent. Their structure arises from statistically induced co-occurrence patterns of AU intensity levels. Modeling this structure is critical for improving the estimation performance; however, this performance is bounded by the quality of the input features extracted from face images. The goal of this paper is to model these structures and estimate complex feature representations simultaneously by combining conditional random field (CRF) encoded AU dependencies with deep learning. To this end, we propose a novel Copula CNN deep learning approach for modeling multivariate ordinal variables. Our model accounts for ordinal structure in output variables and their non-linear dependencies via copula functions modeled as cliques of a CRF. These are jointly optimized with deep CNN feature encoding layers using a newly introduced balanced batch iterative training algorithm. We demonstrate the effectiveness of our approach on the task of AU intensity estimation on two benchmark datasets. We show that joint learning of the deep features and the target output structure results in significant performance gains compared to existing structured deep models and deep models for analysis of facial expressions.

Free-hand sketch-based image retrieval (SBIR) is a specific cross-view retrieval task, in which queries are abstract and ambiguous sketches while the retrieval database is formed with natural images. Work in this area mainly focuses on extracting representative and shared features for sketches and natural images. However, these can neither cope well with the geometric distortion between sketches and images nor be feasible for large-scale SBIR due to the heavy continuous-valued distance computation. In this paper, we speed up SBIR by introducing a novel binary coding method, named Deep Sketch Hashing (DSH), where a semi-heterogeneous deep architecture is proposed and incorporated into an end-to-end binary coding framework. Specifically, three convolutional neural networks are utilized to encode free-hand sketches, natural images and, especially, the auxiliary sketch-tokens which are adopted as bridges to mitigate the sketch-image geometric distortion. The learned DSH codes can effectively capture the cross-view similarities as well as the intrinsic semantic correlations between different categories. To the best of our knowledge, DSH is the first hashing work specifically designed for category-level SBIR with an end-to-end deep architecture. The proposed DSH is comprehensively evaluated on two large-scale datasets of TU-Berlin Extension and Sketchy, and the experiments consistently show DSH's superior SBIR accuracies over several state-of-the-art methods, while achieving significantly reduced retrieval time and memory footprint.

This paper proposes efficient and powerful deep networks for action prediction from partially observed videos containing temporally incomplete action executions. Different from after-the-fact action recognition, action prediction task requires action labels to be predicted from these partially observed videos. Our approach exploits abundant sequential context information to enrich the feature representations of partial videos. We reconstruct missing information in the features extracted from partial videos by learning from fully observed action videos. The amount of the information is temporally ordered for the purpose of modeling temporal orderings of action segments. Label information is also used to better separate the learned features of different categories. We develop a new learning formulation that enables efficient model training. Extensive experimental results on UCF101, Sports-1M and BIT datasets demonstrate that our approach remarkably outperforms state-of-the-art methods, and is up to 300x faster than these methods. Results also show that actions differ in their prediction characteristics; some actions can be correctly predicted even though only the beginning 10% portion of videos is observed.

Estimating dense visual correspondences between objects with intra-class variation, deformations and background clutter remains a challenging problem. Thanks to the breakthrough of CNNs there are new powerful features available. Despite their easy accessibility and great success, existing semantic flow methods could not significantly benefit from these without extensive additional training. We introduce a novel method for semantic matching with pre-trained CNN features which is based on convolutional feature pyramids and activation guided feature selection. For the final matching we propose a sparse graph matching framework where each salient feature selects among a small subset of nearest neighbors in the target image. To improve our method in the unconstrained setting without bounding box annotations we introduce novel object proposal based matching constraints. Furthermore, we show that the sparse matching can be transformed into a dense correspondence field. Extensive experimental evaluations on benchmark datasets show that our method significantly outperforms existing semantic matching methods.

Most existing weakly supervised localization (WSL) approaches learn detectors by finding positive bounding boxes based on features learned with image-level supervision. However, those features do not contain spatial location related information and usually provide poor-quality positive samples for training a detector. To overcome this issue, we propose a deep self-taught learning approach, which makes the detector learn the object-level features reliable for acquiring tight positive samples and afterwards re-train itself based on them. Consequently, the detector progressively improves its detection ability and localizes more informative positive samples. To implement such self-taught learning, we propose a seed sample acquisition method via image-to-object transferring and dense subgraph discovery to find reliable positive samples for initializing the detector. An online supportive sample harvesting scheme is further proposed to dynamically select the most confident tight positive samples and train the detector in a mutual boosting way. To prevent the detector from being trapped in poor optima due to overfitting, we propose a new relative improvement of predicted CNN scores for guiding the self-taught learning process. Extensive experiments on PASCAL 2007 and 2012 show that our approach outperforms the state-of-the-arts, strongly validating its effectiveness.

We propose a new method for creating computationally efficient and compact convolutional neural networks (CNNs) using a novel sparse connection structure that resembles a tree root. This allows a significant reduction in computational cost and number of parameters compared to state-of-the-art deep CNNs, without compromising accuracy, by exploiting the sparsity of inter-layer filter dependencies. We validate our approach by using it to train more efficient variants of state-of-the-art CNN architectures, evaluated on the CIFAR10 and ILSVRC datasets. Our results show similar or higher accuracy than the baseline architectures with much less computation, as measured by CPU and GPU timings. For example, for ResNet 50, our model has 40% fewer parameters, 45% fewer floating point operations, and is 31% (12%) faster on a CPU (GPU). For the deeper ResNet 200 our model has 48% fewer parameters and 27% fewer floating point operations, while maintaining state-of-the-art accuracy. For GoogLeNet, our model has 7% fewer parameters and is 21% (16%) faster on a CPU (GPU).

Generative models of 3D human motion are often restricted to a small number of activities and can therefore not generalize well to novel movements or applications. In this work we propose a deep learning framework for human motion capture data that learns a generic representation from a large corpus of motion capture data and generalizes well to new, unseen, motions. Using an encoding-decoding network that learns to predict future 3D poses from the most recent past, we extract a feature representation of human motion. Most work on deep learning for sequence prediction focuses on video and speech. Since skeletal data has a different structure, we present and evaluate different network architectures that make different assumptions about time dependencies and limb correlations. To quantify the learned features, we use the output of different layers for action classification and visualize the receptive fields of the network units. Our method outperforms the recent state of the art in skeletal motion prediction even though these use action specific training data. Our results show that deep feedforward networks, trained from a generic mocap database, can successfully be used for feature extraction from human motion data and that this representation can be used as a foundation for classification and prediction.

Image captioning is a challenging problem owing to the complexity in understanding the image content and diverse ways of describing it in natural language. Recent advances in deep neural networks have substantially improved the performance of this task. Most state-of-the-art approaches follow an encoder-decoder framework, which generates captions using a sequential recurrent prediction model. However, in this paper, we introduce a novel decision-making framework for image captioning. We utilize a "policy network" and a "value network" to collaboratively generate captions. The policy network serves as a local guidance by providing the confidence of predicting the next word according to the current state. Additionally, the value network serves as a global and lookahead guidance by evaluating all possible extensions of the current state. In essence, it adjusts the goal of predicting the correct words towards the goal of generating captions similar to the ground truth captions. We train both networks using an actor-critic reinforcement learning model, with a novel reward defined by visual-semantic embedding. Extensive experiments and analyses on the Microsoft COCO dataset show that the proposed framework outperforms state-of-the-art approaches across different evaluation metrics.

Deep convolutional neural networks (CNNs) have proven highly effective for visual recognition, where learning a universal representation from activations of convolutional layer plays a fundamental problem. In this paper, we present Fisher Vector encoding with Variational Auto-Encoder (FV-VAE), a novel deep architecture that quantizes the local activations of convolutional layer in a deep generative model, by training them in an end-to-end manner. To incorporate FV encoding strategy into deep generative models, we introduce Variational Auto-Encoder model, which steers a variational inference and learning in a neural network which can be straightforwardly optimized using standard stochastic gradient method. Different from the FV characterized by conventional generative models (e.g., Gaussian Mixture Model) which parsimoniously fit a discrete mixture model to data distribution, the proposed FV-VAE is more flexible to represent the natural property of data for better generalization. Extensive experiments are conducted on three public datasets, i.e., UCF101, ActivityNet, and CUB-200-2011 in the context of video action recognition and fine-grained image classification, respectively. Superior results are reported when compared to state-of-the-art representations. Most remarkably, our proposed FV-VAE achieves to-date the best published accuracy of 94.2% on UCF101.