CVPR 2017 Papers

We propose a unified formulation for the problem of 3D human pose estimation from a single raw RGB image that reasons jointly about 2D joint estimation and 3D pose reconstruction to improve both tasks. We take an integrated approach that fuses probabilistic knowledge of 3D human pose with a multi-stage CNN architecture and uses the knowledge of plausible 3D landmark locations to refine the search for better 2D locations. The entire process is trained end-to-end, is extremely efficient and obtains state-of-the-art results on Human3.6M outperforming previous approaches both on 2D and 3D errors.

Face recognition has the perception of a solved problem, however when tested at the million-scale exhibits dramatic variation in accuracies across the different algorithms [??]. Are the algorithms very different? Is access to good/big training data their secret weapon? Where should face recognition improve? To address those questions, we created a benchmark, MF2, that requires all algorithms to be trained on same data, and tested at the million scale. MF2 is a public large-scale set with 672K identities and 4.7M photos created with the goal to level playing field for large scale face recognition. We contrast our results with findings from the other two large-scale benchmarks MegaFace Challenge and MS-Celebs-1M where groups were allowed to train on any private/public/big/small set. Some key discoveries: 1) algorithms, trained on MF2, were able to achieve state of the art and comparable results to algorithms trained on massive private sets, 2) some outperformed themselves once trained on MF2, 3) invariance to aging suffers from low accuracies as in MegaFace, identifying the need for larger age variations possibly within identities or adjustment of algorithms in future testing.

Image retargeting techniques that adjust images into different sizes have attracted much attention recently. Objective quality assessment (OQA) of image retargeting results is often desired to automatically select the best results. Existing OQA methods output an absolute score for each retargeted image and use these scores to compare different results. Observing that it is challenging even for human subjects to give consistent scores for retargeting results of different source images, in this paper we propose a learning-based OQA method that predicts the ranking of a set of retargeted images with the same source image. We show that this more manageable task helps achieve more consistent prediction to human preference and is sufficient for most application scenarios. To compute the ranking, we propose a simple yet efficient machine learning framework that uses a General Regression Neural Network (GRNN) to model a combination of seven elaborate OQA metrics. We then propose a simple scheme to transform the relative scores output from GRNN into a global ranking. We train our GRNN model using human preference data collected in the elaborate RetargetMe benchmark and evaluate our method based on the subjective study in RetargetMe. Moreover, we introduce a further subjective benchmark to evaluate the generalizability of different OQA methods. Experimental results demonstrate that our method outperforms eight representative OQA methods in ranking prediction and has better generalizability to different datasets.

Confidence measures estimate unreliable disparity assignments performed by a stereo matching algorithm and, as recently proved, can be used for several purposes. This paper aims at increasing, by means of a deep network, the effectiveness of state-of-the-art confidence measures exploiting the local consistency assumption. We exhaustively evaluated our proposal on 23 confidence measures, including 5 top-performing ones based on random-forests and CNNs, training our networks with two popular stereo algorithms and a small subset (25 out of 194 frames) of the KITTI 2012 dataset. Experimental results show that our approach dramatically increases the effectiveness of all the 23 confidence measures on the remaining frames. Moreover, without re-training, we report a further cross-evaluation on KITTI 2015 and Middlebury 2014 confirming that our proposal provides remarkable improvements for each confidence measure even when dealing with significantly different input data. To the best of our knowledge, this is the first method to move beyond conventional pixel-wise confidence estimation.

Understanding the simultaneously very diverse and intricately fine-grained set of possible human actions is a critical open problem in computer vision. Manually labeling training videos is feasible for some action classes but doesn't scale to the full long-tailed distribution of actions. A promising way to address this is to leverage noisy data from web queries to learn new actions, using semi-supervised or "webly-supervised" approaches. However, these methods typically do not learn domain-specific knowledge, or rely on iterative hand-tuned data labeling policies. In this work, we instead propose a reinforcement learning-based formulation for selecting the right examples for training a classifier from noisy web search results. Our method uses Q-learning to learn a data labeling policy on a small labeled training dataset, and then uses this to automatically label noisy web data for new visual concepts. Experiments on the challenging Sports-1M action recognition benchmark as well as on additional fine-grained and newly emerging action classes demonstrate that our method is able to learn good labeling policies for noisy data and use this to learn accurate visual concept classifiers.

We present an end-to-end, multimodal, fully convolutional network for extracting semantic structures from document images. We consider document semantic structure extraction as a pixel-wise segmentation task, and propose a unified model that classifies pixels based not only on their visual appearance, as in the traditional page segmentation task, but also on the content of underlying text. Moreover, we propose an efficient synthetic document generation process that we use to generate pretraining data for our network. Once the network is trained on a large set of synthetic documents, we fine-tune the network on unlabeled real documents using a semi-supervised approach. We systematically study the optimum network architecture and show that both our multimodal approach and the synthetic data pretraining significantly boost the performance.

Deep Neural Networks (DNNs) have substantially improved the state-of-the-art in salient object detection. However, training DNNs requires costly pixel-level annotations. In this paper, we leverage the observation that image-level tags provide important cues of foreground salient objects, and develop a weakly supervised learning method for saliency detection using image-level tags only. The Foreground Inference Network (FIN) is introduced for this challenging task. In the first stage of our training method, FIN is jointly trained with a fully convolutional network (FCN) for image-level tag prediction. A global smooth pooling layer is proposed, enabling FCN to assign object category tags to corresponding object regions, while FIN is capable of capturing all potential foreground regions with the predicted saliency maps. In the second stage, FIN is fine-tuned with its predicted saliency maps as ground truth. For refinement of ground truth, an iterative Conditional Random Field is developed to enforce spatial label consistency and further boost performance. Our method alleviates annotation efforts and allows the usage of existing large scale training sets with image-level tags. Our model runs at 60 FPS, outperforms unsupervised ones with a large margin, and achieves comparable or even superior performance than fully supervised counterparts.

We present an efficient method for geolocalization in urban environments starting from a coarse estimate of the location provided by a GPS and using a simple untextured 2.5D model of the surrounding buildings. Our key contribution is a novel efficient and robust method to optimize the pose: We train a Deep Network to predict the best direction to improve a pose estimate, given a semantic segmentation of the input image and a rendering of the buildings from this estimate. We then iteratively apply this CNN until converging to a good pose. This approach avoids the use of reference images of the surroundings, which are difficult to acquire and match, while 2.5D models are broadly available. We can therefore apply it to places unseen during training.

Statistical decomposition methods are of paramount importance in discovering the modes of variations of visual data. Probably the most prominent linear decomposition method is the Principal Component Analysis (PCA), which discovers a single mode of variation in the data. However, in practice, visual data exhibit several modes of variations. For instance, the appearance of faces varies in identity, expression, pose etc. To extract these modes of variations from visual data, several supervised methods, such as the TensorFaces, that rely on multilinear (tensor) decomposition (e.g., Higher Order SVD) have been developed. The main drawbacks of such methods is that they require both labels regarding the modes of variations and the same number of samples under all modes of variations (e.g., the same face under different expressions, poses etc.). Therefore, their applicability is limited to well-organised data, usually captured in well-controlled conditions. In this paper, we propose the first general multilinear method, to the best of our knowledge, that discovers the multilinear structure of visual data in unsupervised setting. That is, without the presence of labels. We demonstrate the applicability of the proposed method in two applications, namely Shape from Shading (SfS) and expression transfer.

In many real-world scenarios, labeled data for a specific machine learning task is costly to obtain. Semi-supervised training methods make use of abundantly available unlabeled data and a smaller number of labeled examples. We propose a new framework for semi-supervised training of deep neural networks inspired by learning in humans. "Associations" are made from embeddings of labeled samples to those of unlabeled ones and back. The optimization schedule encourages correct association cycles that end up at the same class from which the association was started and penalizes wrong associations ending at a different class. The implementation is easy to use and can be added to any existing end-to-end training setup. We demonstrate the capabilities of learning by association on several data sets and show that it can improve performance on classification tasks tremendously by making use of additionally available unlabeled data. In particular, for cases with few labeled data, our training scheme outperforms the current state of the art on SVHN.

Existing RNN-based approaches for action recognition from depth sequences require either skeleton joints or hand-crafted depth features as inputs. An end-to-end manner, mapping from raw depth maps to action classes, is non-trivial to design due to the fact that: 1) single channel map lacks texture thus weakens the discriminative power; 2) relatively small set of depth training data. To address these challenges, we propose to learn an RNN driven by privileged information (PI) in three-steps: An encoder is pre-trained to learn a joint embedding of depth appearance and PI (i.e. skeleton joints). The learned embedding layers are then tuned in the learning step, aiming to optimize the network by exploiting PI in a form of multi-task loss. However, exploiting PI as a secondary task provides little help to improve the performance of a primary task (i.e. classification) due to the gap between them. Finally, a bridging matrix is defined to connect two tasks by discovering latent PI in the refining step. Our PI-based classification loss maintains a consistency between latent PI and predicted distribution. The latent PI and network are iteratively estimated and updated in an expectation-maximization procedure. The proposed learning process provides greater discriminative power to model subtle depth difference, while helping avoid overfitting the scarcer training data. Our experiments show significant performance gains over state-of-the-art methods on three public benchmark datasets and our newly collected Blanket dataset.

This paper introduces a learning scheme to construct a Hilbert space (i.e., a vector space along its inner product) to address both unsupervised and semi-supervised domain adaptation problems. This is achieved by learning projections from each domain to a latent space along the Mahalanobis metric of the latent space to simultaneously minimizing a notion of domain variance while maximizing a measure of discriminatory power. In particular, we make use of the Riemannian optimization techniques to match statistical properties (e.g., first and second order statistics) between samples projected into the latent space from different domains. Upon availability of class labels, we further deem samples sharing the same label to form more compact clusters while pulling away samples coming from different classes.We extensively evaluate and contrast our proposal against state-of-the-art methods for the task of visual domain adaptation using both handcrafted and deep-net features. Our experiments show that even with a simple nearest neighbor classifier, the proposed method can outperform several state-of-the-art methods benefitting from more involved classification schemes.

Zero-shot learning (ZSL) models rely on learning a joint embedding space where both textual/semantic description of object classes and visual representation of object images can be projected to for nearest neighbour search. Despite the success of deep neural networks that learn an end-to-end model between text and images in other vision problems such as image captioning, very few deep ZSL model exists and they show little advantage over ZSL models that utilise deep feature representations but do not learn an end-to-end embedding. In this paper we argue that the key to make deep ZSL models succeed is to choose the right embedding space. Instead of embedding into a semantic space or an intermediate space, we propose to use the visual space as the embedding space. This is because that in this space, the subsequent nearest neighbour search would suffer much less from the hubness problem and thus become more effective. This model design also provides a natural mechanism for multiple semantic modalities (e.g., attributes and sentence descriptions) to be fused and optimised jointly in an end-to-end manner. Extensive experiments on four benchmarks show that our model significantly outperforms the existing models.

Inspired by recent advances of deep learning in instance segmentation and object tracking, we introduce the concept of convnet-based guidance applied to video object segmentation. Our model proceeds on a per-frame basis, guided by the output of the previous frame towards the object of interest in the next frame. We demonstrate that highly accurate object segmentation in videos can be enabled by using a convolutional neural network (convnet) trained with static images only. The key component of our approach is a combination of offline and online learning strategies, where the former produces a refined mask from the previous' frame estimate and the latter allows to capture the appearance of the specific object instance. Our method can handle different types of input annotations such as bounding boxes and segments while leveraging an arbitrary amount of annotated frames. Therefore our system is suitable for diverse applications with different requirements in terms of accuracy and efficiency. In our extensive evaluation, we obtain competitive results on three different datasets, independently from the type of input annotation.

Multi-label image classification is a fundamental but challenging task in computer vision. Great progress has been achieved by exploiting semantic relations between labels in recent years. However, conventional approaches are unable to model the underlying spatial relations between labels in multi-label images, because spatial annotations of the labels are generally not provided. In this paper, we propose a unified deep neural network that exploits both semantic and spatial relations between labels with only image-level supervisions. Given a multi-label image, our proposed Spatial Regularization Network (SRN) generates attention maps for all labels and captures the underlying relations between them via learnable convolutions. By aggregating the regularized classification results with original results by a ResNet-101 network, the classification performance can be consistently improved. The whole deep neural network is trained end-to-end with only image-level annotations, thus requires no additional efforts on image annotations. Extensive evaluations on 3 public datasets with different types of labels show that our approach significantly outperforms state-of-the-arts and has strong generalization capability. Analysis of the learned SRN model demonstrates that it can effectively capture both semantic and spatial relations of labels for improving classification performance.

We present a learning framework for abstracting complex shapes by learning to assemble objects using 3D volumetric primitives. In addition to generating simple and geometrically interpretable explanations of 3D objects, our framework also allows us to automatically discover and exploit consistent structure in the data. We demonstrate that using our method allows predicting shape representations which can be leveraged for obtaining a consistent parsing across the instances of a shape collection and constructing an interpretable shape similarity measure. We also examine applications for image-based prediction as well as shape manipulation.

Face attributes are interesting due to their detailed description of human faces. Unlike prior researches working on attribute prediction, we address an inverse and more challenging problem called face attribute manipulation which aims at modifying a face image according to a given attribute value. Instead of manipulating the whole image, we propose to learn the corresponding residual image defined as the difference between images before and after the manipulation. In this way, the manipulation can be operated efficiently with modest pixel modification. The framework of our approach is based on the Generative Adversarial Network. It consists of two image transformation networks and a discriminative network. The transformation networks are responsible for the attribute manipulation and its dual operation and the discriminative network is used to distinguish the generated images from real images. We also apply dual learning to allow transformation networks to learn from each other. Experiments show that residual images can be effectively learned and used for attribute manipulations. The generated images remain most of the details in attribute-irrelevant areas.

Large-scale training for semantic segmentation is challenging due to the expense of obtaining training data for this task relative to other vision tasks. We propose a novel training approach to address this difficulty. Given cheaply-obtained sparse image labelings, we propagate the sparse labels to produce guessed dense labelings. A standard CNN-based segmentation network is trained to mimic these labelings. The label-propagation process is defined via random-walk hitting probabilities, which leads to a differentiable parameterization with uncertainty estimates that are incorporated into our loss. We show that by learning the label-propagator jointly with the segmentation predictor, we are able to effectively learn semantic edges given no direct edge supervision. Experiments also show that training a segmentation network in this way outperforms the naive approach.

Recent advanced deep convolutional networks (CNNs) achieved great successes in many computer vision tasks, because of their compelling learning complexity and the presences of large-scale labeled data. However, as obtaining per-pixel annotations is expensive, performances of CNNs in semantic image segmentation are not fully exploited. This work significantly increases segmentation accuracy of CNNs by learning from an Image Descriptions in the Wild (IDW) dataset. Unlike previous image captioning datasets, where captions were manually and densely annotated, images and their descriptions in IDW are automatically downloaded from Internet without any manual cleaning and refinement. An IDW-CNN is proposed to jointly train IDW and existing image segmentation dataset such as Pascal VOC 2012 (VOC). It has two appealing properties. First, knowledge from different datasets can be fully explored and transferred from each other to improve performance. Second, segmentation accuracy in VOC can be constantly increased when selecting more data from IDW. Extensive experiments demonstrate the effectiveness and scalability of IDW-CNN, which outperforms existing best-performing system by 12% on VOC12 test set.