CVPR 2016 Papers

We propose a novel weakly-supervised semantic segmentation algorithm based on Deep Convolutional Neural Net- work (DCNN). Contrary to existing weakly-supervised approaches, our algorithm exploits auxiliary segmentation an- notations available for different categories to guide segmentations on images with only image-level class labels. To make segmentation knowledge transferrable across categories, we design a decoupled encoder-decoder architecture with attention model. In this architecture, the model generates spatial highlights of each category presented in images using an attention model, and subsequently per- forms binary segmentation for each highlighted region using decoder. Combining attention model, the decoder trained with segmentation annotations in different categories boosts accuracy of weakly-supervised semantic segmentation. The proposed algorithm demonstrates substantially improved performance compared to the state-of-the- art weakly-supervised techniques in PASCAL VOC 2012 dataset when our model is trained with the annotations in 60 exclusive categories in Microsoft COCO dataset.

Recent innovations in training deep convolutional neural network (ConvNet) models have motivated the design of new methods to automatically learn local image descriptors. The latest deep ConvNets proposed for this task consist of a siamese network that is trained by penalising misclassification of pairs of local image patches. Current results from machine learning show that replacing this siamese by a triplet network can improve the classification accuracy in several problems, but this has yet to be demonstrated for local image descriptor learning. Moreover, current siamese and triplet networks have been trained with stochastic gradient descent that computes the gradient from individual pairs or triplets of local image patches, which can make them prone to overfitting. In this paper, we first propose the use of triplet networks for the problem of local image descriptor learning. Furthermore, we also propose the use of a global loss that minimises the overall classification error of all patches present in the training set, which can improve the generalisation capability of the model. Using the UBC benchmark dataset for comparing local image descriptors, we show that the triplet network produces a more accurate embedding than the siamese network in terms of the UBC dataset errors. Moreover, we also demonstrate that a combination of the triplet and global losses produces the best embedding in the field, using this triplet network. Finally, we also show that the use of the central-surround siamese network trained with the global loss produces the best result of the field on the UBC dataset.

Can we learn about object classes in images by looking at a collection of relevant 3D models? Or if we want to learn about human (inter-)actions in images, can we benefit from videos or abstract illustrations that show these actions? A common aspect of these settings is the availability of additional or privileged data that can be exploited at training time and that will not be available and not of interest at test time. We seek to generalize the learning with privileged information (LUPI) framework, which requires additional information to be defined per image, to the setting where additional information is a data collection about the task of interest. Our framework minimises the distribution mismatch between errors made in images and in privileged data. The proposed method is tested on four publicly available datasets: Image+ClipArt, Image+3Dobject, and Image+Video. Experimental results reveal that our new LUPI paradigm naturally addresses the cross-dataset learning.

State-of-the-art methods for zero-shot visual recognition formulate learning as a joint embedding problem of images and side information. In these formulations the current best complement to visual features are attributes: manually-encoded vectors describing shared characteristics among categories. Despite good performance, attributes have limitations: (1) finer-grained recognition requires commensurately more attributes, and (2) attributes do not provide a natural language interface. We propose to overcome these limitations by training neural language models from scratch; i.e. without pre-training and only consuming words and characters. Our proposed models train end-to-end to align with the fine-grained and category-specific content of images. Natural language provides a flexible and compact way of encoding only the salient visual aspects for distinguishing categories. By training on raw text, our model can do inference on raw text as well, providing humans a familiar mode both for annotation and retrieval. Our model achieves strong performance on zero-shot text-based image retrieval and significantly outperforms the attribute-based state-of-the-art for zero-shot classification on the Caltech-UCSD Birds 200-2011 dataset.

Data in vision domain often exhibit highly-skewed class distribution, i.e., most data belong to a few majority classes, while the minority classes only contain a scarce amount of instances. To mitigate this issue, contemporary classification methods based on deep convolutional neural network (CNN) typically follow classic strategies such as class re-sampling or cost-sensitive training. In this paper, we conduct extensive and systematic experiments to validate the effectiveness of these classic schemes for representation learning on class-imbalanced data. We further demonstrate that more discriminative deep representation can be learned by enforcing a deep network to maintain both inter-cluster and inter-class margins. This tighter constraint effectively reduces the class imbalance inherent in the local data neighborhood. We show that the margins can be easily deployed in standard deep learning framework through quintuplet instance sampling and the associated triple-header hinge loss. The representation learned by our approach, when combined with a simple k-nearest neighbor (kNN) algorithm, shows significant improvements over existing methods on both high- and low-level vision classification tasks that exhibit imbalanced class distribution.

Attributes possess appealing properties and benefit many computer vision problems, such as object recognition, learning with humans in the loop, and image retrieval. Whereas the existing work mainly pursues utilizing attributes for various computer vision problems, we contend that the most basic problem---how to accurately and robustly detect attributes from images---has been left under explored. Especially, the existing work rarely explicitly tackles the need that attribute detectors should generalize well across different categories, including those previously unseen. Noting that this is analogous to the objective of multi-source domain generalization, if we treat each category as a domain, we provide a novel perspective to attribute detection and propose to gear the techniques in multi-source domain generalization for the purpose of learning cross-category generalizable attribute detectors. We validate our understanding and approach with extensive experiments on four challenging datasets and three different problems.

This paper addresses the challenging problem of perceiving the hidden or occluded geometry of the scene depicted in any given RGBD image. Unlike other image labeling problems such as image segmentation where each pixel needs to be assigned a single label, layered decomposition requires us to assign multiple labels to pixels. We propose a novel "Occlusion-CRF" model that allows for the integration of sophisticated priors to regularize the solution space and enables the automatic inference of the layer decomposition. We use a generalization of the Fusion Move algorithm to perform Maximum a Posterior (MAP) inference on the model that can handle the large label sets needed to represent multiple surface assignments to each pixel. We have evaluated the proposed model and the inference algorithm on many RGBD images of cluttered indoor scenes. Our experiments show that not only is our model able to explain occlusions but it also enables automatic inpainting of occluded/invisible surfaces.

A major goal of imaging studies such as the (ongoing) Human Connectome Project (HCP) is to characterize the structural network map of the human brain and identify its associations with covariates such as genotype, risk factors, and so on that correspond to an individual. But the set of image derived measures and the set of covariates are both large, so we must first estimate a 'parsimonious' set of relations between the measurements. For instance, a Gaussian graphical model will show conditional independences between the random variables, which can then be used to setup specific hypothesis based analyses downstream. But most such data involve a large list of 'latent' variables that remain unobserved, yet affect the 'observed' variables sustantially. Accounting for such latent variables falls outside the scope of standard inverse covariance matrix estimation, and is tackled via highly specialized optimization methods. This paper offers a unique harmonic analysis view of this problem. By casting the estimation of the precision matrix in terms of a composition of low-frequency latent variables and high-frequency sparse terms, we show how the problem can be formulated using a new wavelet-type expansion in non-Euclidean spaces. Our formalization poses the estimation problem entirely in the frequency space and shows how it can be solved by a simple sub-gradient scheme (involving a single variable). We provide a compelling set of scientific results on 500 scans from the recently released HCP data where our algorithm recovers highly interpretable and sparse conditional dependencies between brain connectivity pathways and well-known covariates.

We present a novel latent embedding model for learning a compatibility function between image and class embeddings, in the context of zero-shot classification. The proposed method augments the state-of-the-art bilinear compatibility model by incorporating latent variables. Instead of learning a single bilinear map, it learns a collection of maps with the selection, of which map to use, being a latent variable for the current image-class pair. We train the model with a ranking based objective function which penalizes incorrect rankings of the true class for a given image. We empirically demonstrate that our model improves the state-of-the-art for various class embeddings consistently on three challenging publicly available datasets for the zero-shot setting. Moreover, our method leads to visually highly interpretable results with clear clusters of different fine-grained object properties that correspond to different latent variable maps.

Visual location recognition is the task of determining the place depicted in a query image from a given database of geo-tagged images. Location recognition is often cast as an image retrieval problem and recent research has almost exclusively focused on improving the chance that a relevant database image is ranked high enough after retrieval. The implicit assumption is that the number of inliers found by spatial verification can be used to distinguish between a related and an unrelated database photo with high precision. In this paper, we show that this assumption does not hold for large datasets due to the appearance of geometric bursts, i.e., sets of visual elements appearing in similar geometric configurations in unrelated database photos. We propose algorithms for detecting and handling geometric bursts. Although conceptually simple, using the proposed weighting schemes dramatically improves the recall that can be achieved when high precision is required compared to the standard re-ranking based on the inlier count. Our approach is easy to implement and can easily be integrated into existing location recognition systems

We study the problem of facial analysis in videos. Our first contribution is a novel weakly supervised learning method that models the video event (pain, expression etc.) as a sequence of automatically mined, discriminative sub-events (eg. neutral face, raising brows, contracting lips). The proposed model is inspired by the recent works on Multiple Instance Learning and latent SVM/HCRF- it extends such frameworks to model the ordinal or temporal aspect in the videos, approximately. We show consistent improvements over relevant competitive baselines on four challenging and publicly available video based facial analysis datasets for prediction of expression, clinical pain and intent in dyadic conversations. In combination with complimentary features, we report state-of-the-art results on these datasets.

We present a multi-view reconstruction method that combines conventional multi-view stereo (MVS) with appearance-based normal prediction, to obtain dense and accurate 3D surface models. Reliable surface normals reconstructed from multi-view correspondence serve as training data for a convolutional neural network (CNN), which predicts continuous normal vectors from raw image patches. By training from known points in the same image, the prediction is specifically tailored to the materials and lighting conditions of the particular scene, as well as to the precise camera viewpoint. It is therefore a lot easier to learn than generic single-view normal estimation. The estimated normal maps, together with the known depth values from MVS, are integrated to dense depth maps, which in turn are fused into a 3D model. Experiments on the DTU dataset show that our method delivers 3D reconstructions with the same accuracy as MVS, but with significantly higher completeness.

Interactive image segmentation is an important problem in computer vision with many applications including image editing, object recognition and image retrieval. Most existing interactive segmentation methods only operate on color images. Until recently, very few works have been proposed to leverage depth information from low-cost sensors to improve interactive segmentation. While these methods achieve better results than color-based methods, they are still limited in either using depth as an additional color channel or simply combining depth with color in a linear way. We propose a novel interactive segmentation algorithm which can incorporate multiple feature cues like color, depth, and normals in an unified graph cut framework to leverage these cues more effectively. A key contribution of our method is that it automatically selects a single cue to be used at each pixel, based on the intuition that only one cue is necessary to determine the segmentation label locally. This is achieved by optimizing over both segmentation labels and cue labels, using terms designed to decide where both the segmentation and label cues should change. Our algorithm thus produces not only the segmentation mask but also a cue label map that indicates where each cue contributes to the final result. Extensive experiments on five large scale RGBD datasets show that our proposed algorithm performs significantly better than both other color-based and RGBD based algorithms in reducing the amount of user inputs as well as increasing segmentation accuracy.

Sparse representation has been introduced to visual tracking by finding the best target candidate with minimal reconstruction error within the particle filter framework. However, most sparse representation based trackers have high computational cost, less than promising tracking performance, and limited feature representation. To deal with the above issues, we propose a novel circulant sparse tracker (CST), which exploits circulant target templates. Because of the circulant structure property, CST has the following advantages: (1) It can refine and reduce particles using circular shifts of target templates. (2) The optimization can be efficiently solved entirely in the Fourier domain. (3) High dimensional features can be embedded into CST to significantly improve tracking performance without sacrificing much computation time. Both qualitative and quantitative evaluations on challenging benchmark sequences demonstrate that CST performs better than all other sparse trackers and favorably against state-of-the-art methods.

Automatically generating a natural language description of an image has attracted interests recently both because of its importance in practical applications and because it connects two major artificial intelligence fields: computer vision and natural language processing. Existing approaches are either top-down, which start from a gist of an image and convert it into words, or bottom-up, which come up with words describing various aspects of an image and then combine them. In this paper, we propose a new algorithm that combines both approaches through a model of semantic attention. Our algorithm learns to selectively attend to semantic concept proposals and fuse them into hidden states and outputs of recurrent neural networks. The selection and fusion form a feedback connecting the top-down and bottom-up computation. We evaluate our algorithm on two public benchmarks: Microsoft COCO and Flickr30K. Experimental results show that our algorithm significantly outperforms the state-of-the-art approaches consistently across different evaluation metrics.

Almost all of the current top-performing object detection networks employ region proposals to guide the search for object instances. State-of-the-art region proposal methods usually need several thousand proposals to get high recall, thus hurting the detection efficiency. Although the latest Region Proposal Network method gets promising detection accuracy with several hundred proposals, it still struggles in small-size object detection and precise localization (e.g., large IoU thresholds), mainly due to the coarseness of its feature maps. In this paper, we present a deep hierarchical network, namely HyperNet, for handling region proposal generation and object detection jointly. Our HyperNet is primarily based on an elaborately designed Hyper Feature which aggregates hierarchical feature maps first and then compresses them into a uniform space. The Hyper Features well incorporate deep but highly semantic, intermediate but really complementary, and shallow but naturally high-resolution features of the image, thus enabling us to construct HyperNet by sharing them both in generating proposals and detecting objects via an end-to-end joint training strategy. For the deep VGG16 model, our method achieves completely leading recall and state-of-the-art object detection accuracy on PASCAL VOC 2007 and 2012 using only 100 proposals per image. It runs with a speed of 5 fps (including all steps) on a GPU, thus having the potential for real-time processing.

Hierarchical feature extractors such as Convolutional Networks (ConvNets) have achieved impressive performance on a variety of classification tasks using purely feedforward processing. Feedforward architectures can learn rich representations of the input space but do not explicitly model dependencies in the output spaces, that are quite structured for tasks such as articulated human pose estimation or object segmentation. Here we propose a framework that expands the expressive power of hierarchical feature extractors to encompass both input and output spaces, by introducing top-down feedback. Instead of directly predicting the outputs in one go, we use a self-correcting model that progressively changes an initial solution by feeding back error predictions, in a process we call Iterative Error Feedback (IEF). IEF shows excellent performance on the task of articulated pose estimation in the challenging MPII and LSP benchmarks, matching the state-of-the-art without requiring ground truth scale annotation.

The emergence of wearable devices such as portable cameras and smart glasses makes it possible to record life logging first-person videos. Browsing such long unstructured videos is time-consuming and tedious. This paper studies the discovery of moments of user's major or special interest (i.e., highlights) in a video, for generating the summarization of first-person videos. Specifically, we propose a novel pairwise deep ranking model that employs deep learning techniques to learn the relationship between highlight and non-highlight video segments. A two-stream network structure by representing video segments from complementary information on appearance of video frames and temporal dynamics across frames is developed for video highlight detection. Given a long personal video, equipped with the highlight detection model, a highlight score is assigned to each segment. The obtained highlight segments are applied for summarization in two ways: video timelapse and video skimming. The former plays the highlight (non-highlight) segments at low (high) speed rates, while the latter assembles the sequence of segments with the highest scores. On 100 hours of first-person videos for 15 unique sports categories, our highlight detection achieves the improvement over the state-of-the-art RankSVM method by 10.5% in terms of accuracy. Moreover, our approaches produce video summary with better quality by a user study from 35 human subjects.

Star-convexity prior is popular for interactive single object segmentation due to its simplicity and amenability to binary graph cut optimization. We propose a more general multi-object segmentation approach. Moreover, each object can be constrained by a more descriptive shape prior, "hedgehog". Each hedgehog shape has its surface normals locally constrained by an arbitrary given vector field, e.g. gradient of the user-scribble distance transform. In contrast to star-convexity, the tightness of our normal constraint can be changed giving better control over allowed shapes. For example, looser constraints, i.e. wider cones of allowed normals, give more relaxed hedgehog shapes. On the other hand, the tightest constraint enforces skeleton consistency with the scribbles. In general, hedgehog shapes are more descriptive than a star, which is only a special case corresponding to a radial vector field and weakest tightness. Our approach has significantly more applications than standard single star-convex segmentation, e.g. in medical data we can separate multiple non-star organs with similar appearances and weak edges. Optimization is done by our modified a-expansion moves shown to be submodular for multi-hedgehog shapes.