CVPR 2017 Papers

Visual narrative is often a combination of explicit information and judicious omissions, relying on the viewer to supply missing details. In comics, most movements in time and space are hidden in the "gutters" between panels. To follow the story, readers logically connect panels together by inferring unseen actions through a process called "closure". While computers can now describe the content of natural images, in this paper we examine whether they can understand the closure-driven narratives conveyed by stylized artwork and dialogue in comic book panels. We collect a dataset, COMICS, that consists of over 1.2 million panels (120 GB) paired with automatic textbox transcriptions. An in-depth analysis of COMICS demonstrates that neither text nor image alone can tell a comic book story, so a computer must understand both modalities to keep up with the plot. We introduce three cloze-style tasks that ask models to predict narrative and character-centric aspects of a panel given n preceding panels as context. Various deep neural architectures underperform human baselines on these tasks, suggesting that COMICS contains fundamental challenges for both vision and language.

This paper combines three contributions to establish a new state-of-the-art in dynamic scene recognition. First, we present a novel ConvNet architecture based on temporal residual units that is fully convolutional in spacetime. Our model augments spatial ResNets with convolutions across time to hierarchically add temporal residuals as the depth of the network increases. Second, existing approaches to video-based recognition are categorized and a baseline of seven previously top performing algorithms is selected for comparative evaluation on dynamic scenes. Third, we introduce a new and challenging video database of dynamic scenes that more than doubles the size of those previously available. This dataset is explicitly split into two subsets of equal size that contain videos with and without camera motion to allow for systematic study of how this variable interacts with the defining dynamics of the scene per se. Our evaluations verify the particular strengths and weaknesses of the baseline algorithms with respect to various scene classes and camera motion parameters. Finally, our temporal ResNet boosts recognition performance and establishes a new state-of-the-art on dynamic scene recognition, as well as on the complementary task of action recognition.

The ability to identify and temporally segment fine-grained human actions throughout a video is crucial for robotics, surveillance, education, and beyond. Typical approaches decouple this problem by first extracting local spatiotemporal features from video frames and then feeding them into a temporal classifier that captures high-level temporal patterns. We describe a class of temporal models, which we call Temporal Convolutional Networks (TCNs), that use a hierarchy of temporal convolutions to perform fine-grained action segmentation or detection. Our Encoder-Decoder TCN uses pooling and upsampling to efficiently capture long-range temporal patterns whereas our Dilated TCN uses dilated convolutions. We show that TCNs are capable of capturing action compositions, segment durations, and long-range dependencies, and are over a magnitude faster to train than competing LSTM-based Recurrent Neural Networks. We apply these models to three challenging fine-grained datasets and show large improvements over the state of the art.

Typical techniques for sequence classification are designed for well-segmented sequences which have been edited to remove noisy or irrelevant parts. Therefore, such methods cannot be easily applied on noisy sequences expected in real-world applications. In this paper, we present the Temporal Attention-Gated Model (TAGM) which integrates ideas from attention models and gated recurrent networks to better deal with noisy or unsegmented sequences. Specifically, we extend the concept of attention model to measure the relevance of each observation (time step) of a sequence. We then use a novel gated recurrent network to learn the hidden representation for the final prediction. An important advantage of our approach is interpretability since the temporal attention weights provide a meaningful value for the salience of each time step in the sequence. We demonstrate the merits of our TAGM approach, both for prediction accuracy and interpretability, on three different tasks: spoken digit recognition, text-based sentiment analysis and visual event recognition.

We address the problem of temporal action localization in videos. We pose action localization as a structured prediction over arbitrary-length temporal windows, where each window is scored as the sum of frame-wise classification scores. Additionally, our model classifies the start, middle, and end of each action as separate components, allowing our system to explicitly model each action's temporal evolution and take advantage of informative temporal dependencies present in that structure. In this framework, we localize actions by searching for the structured maximal sum, a problem for which we develop a novel, provably-efficient algorithmic solution. The frame-wise classification scores are computed using features from a deep Convolutional Neural Network (CNN), which are trained end-to-end to directly optimize for a novel structured objective. We evaluate our system on the THUMOS '14 action detection benchmark and achieve competitive performance.

Given two action sequences, we are interested in spotting/co-segmenting all pairs of sub-sequences that represent the same action. We propose a totally unsupervised solution to this problem. No a-priori model of the actions is assumed to be available. The number of common sub-sequences may be unknown. The sub-sequences can be located anywhere in the original sequences, may differ in duration and the corresponding actions may be performed by a different person, in different style. We treat this type of temporal action co-segmentation as a stochastic optimization problem that is solved by employing Particle Swarm Optimization (PSO). The objective function that is minimized by PSO capitalizes on Dynamic Time Warping (DTW) to compare two action sub-sequences. Due to the generic problem formulation and solution, the proposed method can be applied to motion capture (i.e., 3D skeletal) data or to conventional RGB videos acquired in the wild. We present extensive quantitative experiments on standard data sets as well as on data sets we introduced in this paper. The obtained results demonstrate that the proposed method achieves a remarkable increase in co-segmentation quality compared to all tested state of the art methods.

We present in this paper an efficient template-based method for 3D recovery of elastic shapes from a fixed monocular camera. By exploiting the object's elasticity, in contrast to isometric methods that use inextensibility constraints, a large range of deformations can be handled. Our method is expressed as a saddle point problem using Lagrangian multipliers resulting in a linear system which unifies both mechanical and optical constraints and integrates Dirichlet boundary conditions, whether they are fixed or free. We experimentally show that no prior knowledge on material properties is needed, which exhibit the generic usability of our method with elastic and inelastic objects with different kinds of materials. Comparisons with existing techniques are conducted on synthetic and real elastic objects with strains ranging from 25% to 130% resulting to low errors.

We propose a novel measure for template matching named Deformable Diversity Similarity -- based on the diversity of feature matches between a target image window and the template. We rely on both local appearance and geometric information that jointly lead to a powerful approach for matching. Our key contribution is a similarity measure, that is robust to complex deformations, significant background clutter, and occlusions. Empirical evaluation on the most up-to-date benchmark shows that our method outperforms the current state-of-the-art in its detection accuracy while improving computational complexity.

Convolutional neural networks (CNNs) have shown great success in computer vision, approaching human-level performance when trained for specific tasks via application- specific loss functions. In this paper, we propose a method for augmenting and training CNNs so that their learned features are compositional. It encourages networks to form representations that disentangle objects from their surroundings and from each other, thereby promoting better generalization. Our method is agnostic to the specific de- tails of the underlying CNN to which it is applied and can in principle be used with any CNN. As we show in our experiments, the learned representations lead to feature activations that are more localized and improve performance over non-compositional baselines in object recognition tasks.

Integrating complementary features from multiple channels is expected to solve the description ambiguity problem in video captioning, whereas inappropriate fusion strategies often harm rather than help the performance. Existing static fusion methods in video captioning such as concatenation and summation cannot attend to appropriate feature channels, thus fail to adaptively support the recognition of various kinds of visual entities such as actions and objects. This paper contributes to: 1)The first in-depth study of the weakness inherent in data-driven static fusion methods for video captioning. 2) The establishment of a task-driven dynamic fusion (TDDF) method. It can adaptively choose different fusion patterns according to model status. 3) The improvement of video captioning. Extensive experiments conducted on two well-known benchmarks demonstrate that our dynamic fusion method outperforms the state-of-the-art results on MSVD with METEOR scores 0.333, and achieves superior METEOR scores 0.278 on MSR-VTT-10K. Compared to single features, the relative improvement derived from our fusion method are 10.0% and 5.7% respectively on two datasets.

Vision and language understanding has emerged as a subject undergoing intense study in Artificial Intelligence. Among many tasks in this line of research, visual question answering (VQA) has been one of the most successful ones, where the goal is to learn a model that understands visual content at region-level details and finds their associations with pairs of questions and answers in the natural language form. Despite the rapid progress in the past few years, most existing work in VQA have focused primarily on images. In this paper, we focus on extending VQA to the video domain and contribute to the literature in three important ways. First, we propose three new tasks designed specifically for video VQA, which require spatio-temporal reasoning from videos to answer questions correctly. Next, we introduce a new large-scale dataset for video VQA named TGIF-QA that extends existing VQA work with our new tasks. Finally, we propose a dual-LSTM based approach with both spatial and temporal attention, and show its effectiveness over conventional VQA techniques through empirical evaluations.

We present a method for synthesizing a frontal, neutral-expression image of a person's face, given an input face photograph. This is achieved by learning to generate facial landmarks and textures from features extracted from a facial-recognition network. Unlike previous generative approaches, our encoding feature vector is largely invariant to lighting, pose, and facial expression. Exploiting this invariance, we train our decoder network using only frontal, neutral-expression photographs. Since these photographs are well aligned, we can decompose them into a sparse set of landmark points and aligned texture maps. The decoder then predicts landmarks and textures independently and combines them using a differentiable image warping operation. The resulting images can be used for a number of applications, such as analyzing facial attributes, exposure and white balance adjustment, or creating a 3-D avatar.

Video sequences contain rich dynamic patterns, such as dynamic texture patterns that exhibit stationarity in the temporal domain, and action patterns that are non-stationary in either spatial or temporal domain. We show that a spatial-temporal generative ConvNet can be used to model and synthesize dynamic patterns. The model defines a probability distribution on the video sequence, and the log probability is defined by a spatial-temporal ConvNet that consists of multiple layers of spatial-temporal filters to capture spatial-temporal patterns of different scales. The model can be learned from the training video sequences by an "analysis by synthesis" learning algorithm that iterates the following two steps. Step 1 synthesizes video sequences from the currently learned model. Step 2 then updates the model parameters based on the difference between the synthesized video sequences and the observed training sequences. We show that the learning algorithm can synthesize realistic dynamic patterns.

We study the problem of learning generative models of 3D shapes. Voxels or 3D parts have been widely used as the underlying representations to build complex 3D shapes; however, voxel-based representations suffer from high memory requirements, and parts-based models require a large collection of cached or richly parametrized parts. We take an alternative approach: learning a generative model over multi-view depth maps or their corresponding silhouettes, and using a deterministic rendering function to produce 3D shapes from these images. A multi-view representation of shapes enables generation of 3D models with fine details, as 2D depth maps and silhouettes can be modeled at a much higher resolution than 3D voxels. Moreover, our approach naturally brings the ability to recover the underlying 3D representation from depth maps of one or a few viewpoints. Experiments show that our framework can generate 3D shapes with variations and details. We also demonstrate that our model has out-of-sample generalization power for real-world tasks with occluded objects.

In this paper, we study the problem of semantic annotation on 3D models that are represented as shape graphs. A functional view is taken to represent localized information on graphs, so that annotations such as part segment or keypoint are nothing but 0-1 indicator vertex functions. Compared with images that are 2D grids, shape graphs are irregular and non-isomorphic data structures. To enable the prediction of vertex functions on them by convolutional neural networks, we resort to spectral CNN method that enables weight sharing by parametrizing kernels in the spectral domain spanned by graph Laplacian eigenbases. Under this setting, our network, named SyncSpecCNN, strives to overcome two key challenges: how to share coefficients and conduct multi-scale analysis in different parts of the graph for a single shape, and how to share information across related but different shapes that may be represented by very different graphs. Towards these goals, we introduce a spectral parametrization of dilated convolutional kernels and a spectral transformer network. Experimentally we tested SyncSpecCNN on various tasks, including 3D shape part segmentation and keypoint prediction. State-of-the-art performance has been achieved on all benchmark datasets.

We propose a novel crowd counting model that maps a given crowd scene to its density. Crowd analysis is compounded by myriad of factors like inter-occlusion between people due to extreme crowding, high similarity of appearance between people and background elements, and large variability of camera view-points. Current state-of-the art approaches tackle these factors by using multi-scale CNN architectures, recurrent networks and late fusion of features from multi-column CNN with different receptive fields. We propose switching convolutional neural network that leverages variation of crowd density within an image to improve the accuracy and localization of the predicted crowd count. Patches from a grid within a crowd scene are relayed to independent CNN regressors based on crowd count prediction quality of the CNN established during training. The independent CNN regressors are designed to have different receptive fields and a switch classifier is trained to relay the crowd scene patch to the best CNN regressor. We perform extensive experiments on all major crowd counting datasets and evidence better performance compared to current state-of-the-art methods. We provide interpretable representations of the multichotomy of space of crowd scene patches inferred from the switch. It is observed that the switch relays an image patch to a particular CNN column based on density of crowd.

Surveillance video parsing, which segments the video frames into several labels, i.e., face, pants, left-leg, has wide applications. However, annotating all frames pixel-wisely is tedious and inefficient. In this paper, we develop a Single frame Video Parsing (SVP) method which requires only one labeled frame per video in training stage. To parse one particular frame, the video segment preceding the frame is jointly considered. SVP 1: roughly parses the frames within the video segment, 2: estimates the optical flow between frames and 3: fuses the rough parsing results warped by optical flow to produce the refined parsing result. The three components of SVP, namely frame parsing, optical flow estimation and temporal fusion are integrated in an end-to-end manner. Experimental results on two surveillance video datasets reveal that SVP is superior than state-of-the-arts.

We address the problem of transferring motion between captured 4D models. We particularly focus on human subjects for which the ability to automatically augment 4D datasets, by propagating movements between subjects, is of interest in a great deal of recent vision applications that builds on human visual corpus. Given 4D training sets for two subjects for which a sparse set of corresponding keyposes are known, our method is able to transfer a newly captured motion from one subject to the other. With the aim to generalize transfers to input motions possibly very diverse with respect to the training sets, the method contributes with a new transfer model based on non-linear pose interpolation. Building on Gaussian process regression, this model intends to capture and preserve individual motion properties, and thereby realism, by accounting for pose inter-dependencies during motion transfers. Our experiments show visually qualitative, and quantitative, improvements over existing pose-mapping methods and confirm the generalization capabilities of our method compared to state of the art.

3D shape models are naturally parameterized using vertices and faces, i.e, composed on polygons forming a surface. However, current 3D learning paradigms for predictive and generative tasks using convolutional neural networks focus on a voxelized representation of the object. Lifting convolution operators from the traditional 2D to 3D results in high computational overhead with little additional benefit as most of the geometry information is contained on the surface boundary. Here we study the problem of directly generating the 3D shape surface of rigid and non-rigid shapes using deep convolutional neural networks. We develop a procedure to create consistent `geometry images' representing the 3D shape surface of a category of shapes. We then use this consistent representation for category-specific shape generation from a parametric representation or an image by developing novel extensions of deep residual networks for the task of 3D surface generation. Our experiments indicate that our network learns a meaningful representation of shape surfaces allowing it to interpolate between shape orientations and poses, invent new shape surfaces, reconstruct 3D shape surfaces from previously unseen images, and rectify noisy correspondence between 3D shapes belonging to the same class.