CVPR 2017 Papers

"If I provide you a face image of mine (without telling you the actual age when I took the picture) and a large amount of face images that I crawled (containing labeled faces of different ages but not necessarily paired), can you show me what I would look like when I am 80 or what I was like when I was 5?" The answer is probably a "No." Most existing face aging works attempt to learn the transformation between age groups and thus would require the paired samples as well as the labeled query image. In this paper, we look at the problem from a generative modeling perspective such that no paired samples is required. In addition, given an unlabeled image, the generative model can directly produce the image with desired age attribute. We propose a conditional adversarial autoencoder (CAAE) that learns a face manifold, traversing on which smooth age progression and regression can be realized simultaneously. In CAAE, the face is first mapped to a latent vector through a convolutional encoder, and then the vector is projected to the face manifold conditional on age through a deconvolutional generator. The latent vector preserves personalized face features (i.e., personality) and the age condition controls progression vs. regression. Two adversarial networks are imposed on the encoder and generator, respectively, forcing to generate more photo-realistic faces. Experimental results demonstrate the appealing performance and flexibility of the proposed framework by comparing with the state-of-the-art and ground truth.

Generalization performance of trained computer vision (CV) systems that use computer graphics (CG) generated data is not yet effective due to the concept of 'domain-shift' between virtual and real data. Although simulated data augmented with a few real-world samples has been shown to mitigate domain shift and improve transferability of trained models, guiding or bootstrapping the virtual data generation with the distributions learnt from target real world domain is desired, especially in the fields where annotating even few real images is laborious (such as semantic labeling, optical flow, and intrinsic images etc.). In order to address this problem in an unsupervised manner, our work combines recent advances in CG, which aims at generating stochastic scene layouts using large collections of 3D object models, and generative adversarial training, which aims at training generative models by measuring discrepancy between generated and real data in terms of their separability in the space of a deep discriminatively-trained classifier. Our method uses iterative estimation of the posterior density of prior distributions for a generative graphical model. This is done within a rejection sampling framework. Initially, we assume uniform distributions as priors over parameters of a scene described by a generative graphical model. As iterations proceed the uniform prior distributions are updated sequentially to distributions that are closer to the unknown distributions of target data. We demonstrate the utility of adversarially tuned scene generation on two real world benchmark datasets (CityScapes and CamVid) for traffic scene semantic labeling with a deep convolutional net (DeepLab). We obtained performance improvements by 2.28 and 3.14 points on the IoU metric between the DeepLab models trained on simulated sets prepared from the scene generation models before and after tuning to CityScapes and CamVid respectively.

Adversarial learning methods are a promising approach to training robust deep networks, and can generate complex samples across diverse domains. They can also improve recognition despite the presence of domain shift or dataset bias: recent adversarial approaches to unsupervised domain adaptation reduce the difference between the training and test domain distributions and thus improve generalization performance. However, while generative adversarial networks (GANs) show compelling visualizations, they are not optimal on discriminative tasks and can be limited to smaller shifts. On the other hand, discriminative approaches can handle larger domain shifts, but impose tied weights on the model and do not exploit a GAN-based loss. In this work, we first outline a novel generalized framework for adversarial adaptation, which subsumes recent state-of-the-art approaches as special cases, and use this generalized view to better relate prior approaches. We then propose a previously unexplored instance of our general framework which combines discriminative modeling, untied weight sharing, and a GAN loss, which we call Adversarial Discriminative Domain Adaptation (ADDA). We show that ADDA is more effective yet considerably simpler than competing domain-adversarial methods, and demonstrate the promise of our approach by exceeding state-of-the-art unsupervised adaptation results on standard domain adaptation tasks as well as a difficult cross-modality object classification task.

Principal component analysis (PCA) is one of the most versatile tools for unsupervised learning with applications ranging from dimensionality reduction to exploratory data analysis and visualization. While much effort has been devoted to encouraging meaningful representations through regularization (e.g. non-negativity or sparsity), underlying linearity assumptions can limit their effectiveness. To address this issue, we propose Additive Component Analysis (ACA), a novel nonlinear extension of PCA. Inspired by multivariate nonparametric regression with additive models, ACA fits a smooth manifold to data by learning an explicit mapping from a low-dimensional latent space to the input space, which trivially enables applications like denoising. Furthermore, ACA can be used as a drop-in replacement in many algorithms that use linear component analysis methods as a subroutine via the local tangent space of the learned manifold. Unlike many other nonlinear dimensionality reduction techniques, ACA can be efficiently applied to large datasets since it does not require computing pairwise similarities or storing training data during testing. Multiple ACA layers can also be composed and learned jointly with essentially the same procedure for improved representational power, demonstrating the encouraging potential of nonparametric deep learning. We evaluate ACA on a variety of datasets, showing improved robustness, reconstruction performance, and interpretability.

Many computer vision problems require optimization of binary non-submodular energies. In this context, iterative submodularization techniques based on trust region (LSA-TR) and auxiliary functions (LSA-AUX) have been recently proposed [??]. They achieve state-of-the-art-results on a number of computer vision applications. In this paper we extend the LSA-AUX framework in two directions. First, unlike LSA-AUX which selects auxiliary functions based solely on the current solution, we propose to incorporate several additional criteria. This results in tighter bounds for configurations that are more likely or closer to the current solution. Second, we propose move-making extensions of LSA-AUX which achieve tighter bounds by restricting the search space. Finally, we evaluate our methods on several applications. We show that for each application at least one of our extensions significantly outperforms the original LSA-AUX. Moreover, the best extension of LSA-AUX is comparable to or better than LSA-TR on five out of six applications, achieving state-of-the-arts results on four out of six applications.

Many modern computer vision and machine learning applications rely on solving difficult optimization problems that involve non-differentiable objective functions and constraints. The alternating direction method of multipliers (ADMM) is a widely used approach to solve such problems. Relaxed ADMM is a generalization of ADMM that often achieves better performance, but its efficiency depends strongly on algorithm parameters that must be chosen by an expert user. We propose an adaptive method that automatically tunes the key algorithm parameters to achieve optimal performance without user oversight. Inspired by recent work on adaptivity, the proposed adaptive relaxed ADMM (ARADMM) is derived by assuming a Barzilai-Borwein style linear gradient. A detailed convergence analysis of ARADMM is provided, and numerical results on several applications demonstrate fast practical convergence.

In linear representation-based image classification, an unlabeled sample is represented by the entire training set. To obtain a stable and discriminative solution, regularization on the vector of representation coefficients is necessary. For example, the representation in sparse representation-based classification (SRC) uses L1 norm penalty as regularization, which is equal to lasso. However, lasso overemphasizes the role of sparseness while ignoring the inherent structure among samples belonging to a same class. Many recent developed representation classifications have adopted lasso-type regressions to improve the performance. In this paper, we propose the adaptive class preserving representation for classification (ACPRC). Our method is related to group lasso based classification but different in two key points: When training samples in a class are uncorrelated, ACPRC turns into SRC; when samples in a class are highly correlated, it obtains similar result as group lasso. The superiority of ACPRC over other state-of-the-art regularization techniques including lasso, group lasso, sparse group lasso, etc. are evaluated by extensive experiments.

We propose a novel method for temporally pooling frames in a video for the task of human action recognition. The method is motivated by the observation that there are only a small number of frames which, together, contain sufficient information to discriminate an action class present in a video, from the rest. The proposed method learns to pool such discriminative and informative frames, while discarding a majority of the non-informative frames in a single temporal scan of the video. Our algorithm does so by continuously predicting the discriminative importance of each video frame and subsequently pooling them in a deep learning framework. We show the effectiveness of our proposed pooling method on standard benchmarks where it consistently improves on baseline pooling methods, with both RGB and optical flow based Convolutional networks. Further, in combination with complementary video representations, we show results that are competitive with respect to the state-of-the-art results on two challenging and publicly available benchmark datasets.

In recent years, deep learning has achieved great success in many computer vision applications. Convolutional neural networks (CNNs) have lately emerged as a major approach to image classification. Most research on CNNs thus far has focused on developing architectures such as the Inception and residual networks. The convolution layer is the core of the CNN, but few studies have addressed the convolution unit itself. In this paper, we introduce a convolution unit called the active convolution unit (ACU). A new convolution has no fixed shape, because of which we can define any form of convolution. Its shape can be learned through backpropagation during training. Our proposed unit has a few advantages. First, the ACU is a generalization of convolution; it can define not only all conventional convolutions, but also convolutions with fractional pixel coordinates. We can freely change the shape of the convolution, which provides greater freedom to form CNN structures. Second, the shape of the convolution is learned while training and there is no need to tune it by hand. Third, the ACU can learn better than a conventional unit, where we obtained the improvement simply by changing the conventional convolution to an ACU. We tested our proposed method on plain and residual networks, and the results showed significant improvement using our method on various datasets and architectures in comparison with the baseline. Code is available at https://github.com/jyh2986/Active-Convolution.

In this work, we introduce a new video representation for action classification that aggregates local convolutional features across the entire spatio-temporal extent of the video. We do so by integrating state-of-the-art two-stream networks with learnable spatio-temporal feature aggregation. The resulting architecture is end-to-end trainable for whole-video classification. We investigate different strategies for pooling across space and time and combining signals from the different streams. We find that: (i) it is important to pool jointly across space and time, but (ii) appearance and motion streams are best aggregated into their own separate representations. Finally, we show that our representation outperforms the two-stream base architecture by a large margin (13% relative) as well as outperforms other baselines with comparable base architectures on HMDB51, UCF101, and Charades video classification benchmarks.

This paper proposes a novel tracker which is controlled by sequentially pursuing actions learned by deep reinforcement learning. In contrast to the existing trackers using deep networks, the proposed tracker is designed to achieve a light computation as well as satisfactory tracking accuracy in both location and scale. The deep network to control actions is pre-trained using various training sequences and fine-tuned during tracking for online adaptation to target and background changes. The pre-training is done by utilizing deep reinforcement learning as well as supervised learning. The use of reinforcement learning enables even partially labeled data to be successfully utilized for semi-supervised learning. Through evaluation of the OTB dataset, the proposed tracker is validated to achieve a competitive performance that is three times faster than state-of-the-art, deep network-based trackers. The fast version of the proposed method, which operates in real-time on GPU, outperforms the state-of-the-art real-time trackers.

Action Unit (AU) detection becomes essential for facial analysis. Many proposed approaches face challenging problems in dealing with the alignments of different face regions, in the effective fusion of temporal information, and in training a model for multiple AU labels. To better address these problems, we propose a deep learning framework for AU detection with region of interest (ROI) adaptation, integrated multi-label learning, and optimal LSTM-based temporal fusing. First, an ROI cropping net is designed to make sure specific interested regions of faces are learned independently; each sub-region has a local convolutional neural network (CNN) whose convolutional filters will only be trained for the corresponding region. Second, multi-label learning is employed to integrate the outputs of those individual ROI cropping nets, which learns the inter-relationships of various AUs and acquires global features across sub-regions for AU detection. Finally, the optimal selection of multiple LSTM layers are carried out to best fuse temporal features, in order to make the AU prediction the most accurate. The proposed approach is evaluated on two popular AU detection datasets, BP4D and DISFA, outperforming the state of the art significantly, with an average improvement of around 13% in BP4D and 25% in DISFA, respectively.

We propose a novel, practical solution for high quality reconstruction of axially-symmetric transparent objects. While a special case, such transparent objects are ubiquitous in the real world. Common examples of these are glasses, goblets, tumblers, carafes, etc., that can have very unique and visually appealing forms making their reconstruction interesting for vision and graphics applications. Our acquisition setup involves imaging such objects from a single viewpoint while illuminating them from directly behind with a few patterns emitted by an LCD panel. Our reconstruction step is then based on optimization of the object's geometry and its refractive index to minimize the difference between observed and simulated transmission/refraction of rays passing through the object. We exploit the object's axial symmetry as a strong shape prior which allows us to achieve robust reconstruction from a single viewpoint using a simple, commodity acquisition setup. We demonstrate high quality reconstruction of several common rotationally symmetric as well as more complex n-fold symmetric transparent objects with our approach.

Most of the recent successful methods in accurate object detection and localization used some variants of R-CNN style two stage Convolutional Neural Networks (CNN) where plausible regions were proposed in the first stage then followed by a second stage for decision refinement. Despite the simplicity of training and the efficiency in deployment, the single stage detection methods have not been as competitive when evaluated in benchmarks consider mAP for high IoU thresholds. In this paper, we proposed a novel single stage end-to-end trainable object detection network to overcome this limitation. We achieved this by introducing Recurrent Rolling Convolution (RRC) architecture over multi-scale feature maps to construct object classifiers and bounding box regressors which are "deep in context". We evaluated our method in the challenging KITTI dataset which measures methods under IoU threshold of 0.7. We showed that with RRC, a single reduced VGG-16 based model already significantly outperformed all the previously published results. At the time this paper was written our models ranked the first in KITTI car detection (the hard level), the first in cyclist detection and the second in pedestrian detection. These results were not reached by the previous single stage methods. The code is publicly available.

We present an optical flow estimation approach that operates on the full four-dimensional cost volume. This direct approach shares the structural benefits of leading stereo matching pipelines, which are known to yield high accuracy. To this day, such approaches have been considered impractical due to the size of the cost volume. We show that the full four-dimensional cost volume can be constructed in a fraction of a second due to its regularity. We then exploit this regularity further by adapting semi-global matching to the four-dimensional setting. This yields a pipeline that achieves significantly higher accuracy than state-of-the-art optical flow methods while being faster than most. Our approach outperforms all published general-purpose optical flow methods on both Sintel and KITTI 2015 benchmarks.

We introduce a novel approach to jointly estimate consistent depth and normal maps from 4D light fields, with two main contributions. First, we build a cost volume from focal stack symmetry. However, in contrast to previous approaches, we introduce partial focal stacks in order to be able to robustly deal with occlusions. This idea already yields significanly better disparity maps. Second, even recent sublabel-accurate methods for multi-label optimization recover only a piecewise flat disparity map from the cost volume, with normals pointing mostly towards the image plane. This renders normal maps recovered from these approaches unsuitable for potential subsequent applications. We therefore propose regularization with a novel prior linking depth to normals, and imposing smoothness of the resulting normal field. We then jointly optimize over depth and normals to achieve estimates for both which surpass previous work in accuracy on a recent benchmark.

Most existing hashing methods resort to binary codes for similarity search, owing to the high efficiency of computation and storage. However, binary codes lack enough capability in similarity preservation, resulting in less desirable performance. To address this issue, we propose an asymmetric multi-valued hashing method supported by two different non-binary embeddings. (1) A real-valued embedding is used for representing the newly-coming query. (2) A multi-integer-embedding is employed for compressing the whole database, which is modeled by binary sparse representation with fixed sparsity. With these two non-binary embeddings, the similarities between data points can be preserved precisely. To perform meaningful asymmetric similarity computation for efficient semantic search, these embeddings are jointly learnt by preserving the label-based similarity. Technically, this results in a mixed integer programming problem, which is efficiently solved by alternative optimization. Extensive experiments on three multilabel datasets demonstrate that our approach not only outperforms the existing binary hashing methods in search accuracy, but also retains their query and storage efficiency.

Given a user's query, traditional image search systems rank images according to its relevance to a single modality (e.g., image content or surrounding text). Nowadays, an increasing number of images on the Internet are available with associated meta data in rich modalities (e.g., titles, keywords, tags, etc.), which can be exploited for better similarity measure with queries. In this paper, we leverage visual and textual modalities for image search by learning their correlation with input query. According to the intent of query, attention mechanism can be introduced to adaptively balance the importance of different modalities. We propose a novel Attention guided Multi-modal Correlation (AMC) learning method which consists of a jointly learned hierarchy of intra and inter-attention networks. Conditioned on query's intent, intra-attention networks (i.e., visual intra-attention network and language intra-attention network) attend on informative parts within each modality; a multi-modal inter-attention network promotes the importance of the most query-relevant modalities. In experiments, we evaluate AMC models on the search logs from two real world image search engines and show a significant boost on the ranking of user-clicked images in search results. Additionally, we extend AMC models to caption ranking task on COCO dataset and achieve competitive results compared with recent state-of-the-arts.

We consider the problem of data augmentation, i.e., generating artificial samples to extend a given corpus of training data. Specifically, we propose attributed-guided augmentation (AGA) which learns a mapping that allows to synthesize data such that an attribute of a synthesized sample is at a desired value or strength. This is particularly interesting in situations where little data with no attribute annotation is available for learning, but we have access to a large external corpus of heavily annotated samples. While prior works primarily augment in the space of images, we propose to perform augmentation in feature space instead. We implement our approach as a deep encoder-decoder architecture that learns the synthesis function in an end-to-end manner. We demonstrate the utility of our approach on the problems of (1) one-shot object recognition in a transfer-learning setting where we have no prior knowledge of the new classes, as well as (2) object-based one-shot scene recognition. As external data, we leverage 3D depth and pose information from the SUN RGB-D dataset. Our experiments show that attribute-guided augmentation of high-level CNN features considerably improves one-shot recognition performance on both problems.