ICCV 2017 Papers

Following the gaze of people inside videos is an important signal for understanding people and their actions. In this paper, we present an approach for following gaze in video by predicting where a person (in the video) is looking even when the object is in a different frame. We collect VideoGaze, a new dataset which we use as a benchmark to both train and evaluate models. Given one frame with a person in it, our model estimates a density for gaze location in every frame and the probability that the person is looking in that particular frame. A key aspect of our approach is an end-to-end model that jointly estimates: saliency, gaze pose, and geometric relationships between views while only using gaze as supervision. Visualizations suggest that the model learns to internally solve these intermediate tasks automatically without additional supervision. Experiments show that our approach follows gaze in video better than existing approaches, enabling a richer understanding of human activities in video.

Scene text recognition has been a hot research topic in computer vision due to its various applications. The state of the art is the attention-based encoder-decoder framework that learns the mapping between input images and output sequences in a purely data-driven way. However, we observe that existing attention-based methods perform poorly on complicated and/or low-quality images. One major reason is that existing methods cannot get accurate alignments between feature areas and targets for such images. We call this phenomenon "attention drift". To tackle this problem, in this paper we propose the FAN (the abbreviation of Focusing Attention Network) method that employs a focusing attention mechanism to automatically draw back the drifted attention. FAN consists of two major components: an attention network (AN) that is responsible for recognizing character targets as in the existing methods, and a focusing network (FN) that is responsible for adjusting attention by evaluating whether AN pays attention properly on the target areas in the images. Furthermore, different from the existing methods, we adopt a ResNet-based network to enrich deep representations of scene text images. Extensive experiments on various benchmarks, including the IIIT5k, SVT and ICDAR datasets, show that the FAN method substantially outperforms the existing methods.

The focal track sensor is a monocular and computationally efficient depth sensor that is based on defocus controlled by a liquid membrane lens. It synchronizes small lens oscillations with a photosensor to produce real-time depth maps by means of differential defocus, and it couples these oscillations with bigger lens deformations that adapt the defocus working range to track objects over large axial distances. To create the focal track sensor, we derive a texture-invariant family of equations that relate image derivatives to scene depth when a lens changes its focal length differentially. Based on these equations, we design a feed-forward sequence of computations that: robustly incorporates image derivatives at multiple scales; produces confidence maps along with depth; and can be trained end-to-end to mitigate against noise, aberrations, and other non-idealities. Our prototype with 1-inch optics produces depth and confidence maps at 100 frames per second over an axial range of more than 75cm.

Extending state-of-the-art object detectors from image to video is challenging. The accuracy of detection suffers from degenerated object appearances in videos, e.g., motion blur, video defocus, rare poses, etc. Existing work attempts to exploit temporal information on box level, but such methods are not trained end-to-end. We present flow-guided feature aggregation, an accurate and end-to-end learning framework for video object detection. It leverages temporal coherence on feature level instead. It improves the per-frame features by aggregation of nearby features along the motion paths, and thus improves the video recognition accuracy. Our method significantly improves upon strong single-frame baselines in ImageNet VID, especially for more challenging fast moving objects. Our framework is principled, and on par with the best engineered systems winning the ImageNet VID challenges 2016, without additional bells-and-whistles.

In action recognition, local motion descriptors contribute to effectively representing video sequences where target actions appear in localized spatio-temporal regions. For robust recognition, those fundamental descriptors are required to be invariant against horizontal (mirror) flipping in video frames which frequently occurs due to changes of camera viewpoints and action directions, deteriorating classification performance. In this paper, we propose methods to render flip invariance to the local motion descriptors by two approaches. One method leverages local motion flows to ensure the invariance on input patches where the descriptors are computed. The other derives a invariant form theoretically from the flipping transformation applied to hand-crafted descriptors. The method is also extended so as to deal with ConvNet descriptors through learning the invariant form based on data. The experimental results on human action classification show that the proposed methods favorably improve performance both of the handcrafted and the ConvNet descriptors.

While fine-grained object recognition is an important problem in computer vision, current models are unlikely to accurately classify objects in the wild. These fully supervised models need additional annotated images to classify objects in every new scenario, a task that is infeasible. However, sources such as e-commerce websites and field guides provide annotated images for many classes. In this work, we study fine-grained domain adaptation as a step towards overcoming the dataset shift between easily acquired annotated images and the real world. Adaptation has not been studied in the fine-grained setting where annotations such as attributes could be used to increase performance. Our work uses an attribute based multitask adaptaion loss to increase accuracy from a baseline of 3.4% to 19% in the semi-supervised adaptation case. Prior domain adaptation works have been benchmarked on small datasets such as [45] with a total of 795 images for some domains, or simplistic datasets such as [40] consisting of digits. We perform experiments on a new challenging fine-grained dataset of cars consisting of 1, 095, 021 images of 2, 657 categories of cars drawn from e-commerce websites and Google Street View.

While recovery of hyperspectral signals from natural RGB images has been a recent subject of exploration, little to no consideration has been given to the camera response profiles used in the recovery process. In this paper we demonstrate that optimal selection of camera response filters may improve hyperspectral estimation accuracy by over 33%, emphasizing the importance of considering and selecting these response profiles wisely. Additionally, we present an evolutionary optimization methodology for optimal filter set selection from very large filter spaces, an approach that facilitates practical selection from families of customizable filters or filter optimization for multispectral cameras with more than 3 channels.

Facial alignment involves finding a set of landmark points on an image with a known semantic meaning. However, this semantic meaning of landmark points is often lost in 2D approaches where landmarks are either moved to visible boundaries or ignored as the pose of the face changes. In order to extract consistent alignment points across large poses, the 3D structure of the face must be considered in the alignment step. However, extracting a 3D structure from a single 2D image usually requires alignment in the first place. We present our novel approach to simultaneously extract the 3D shape of the face and the semantically consistent 2D alignment through a 3D Spatial Transformer Network (3DSTN) to model both the camera projection matrix and the warping parameters of a 3D model. By utilizing a generic 3D model and a Thin Plate Spline (TPS) warping function, we are able to generate subject specific 3D shapes without the need for a large 3D shape basis. In addition, our proposed network can be trained in an end-to-end framework on entirely synthetic data from the 300W-LP dataset. Unlike other 3D methods, our approach only requires one pass through the network resulting in a faster than real-time alignment. Evaluations of our model on the Annotated Facial Landmarks in the Wild (AFLW) and AFLW2000-3D datasets show our method achieves state-of-the-art performance over other 3D approaches to alignment.

We consider the problem of finding consistent matches across multiple images. Current state-of-the-art solutions use constraints on cycles of matches together with convex optimization, leading to computationally intensive iterative algorithms. In this paper, we instead propose a clustering-based formulation: we first rigorously show its equivalence with traditional approaches, and then propose QuickMatch, a novel algorithm that identifies multi-image matches from a density function in feature space. Specifically, QuickMatch uses the density estimate to order the points in a tree, and then extracts the matches by breaking this tree using feature distances and measures of distinctiveness. Our algorithm outperforms previous state-of-the-art methods (such as MatchALS) in accuracy, and it is significantly faster (up to 62 times faster on some benchmarks), and can scale to large datasets (with more than twenty thousands features).

We present an approach to accelerating a wide variety of image processing operators. Our approach uses a fully-convolutional network that is trained on input-output pairs that demonstrate the operator's action. After training, the original operator need not be run at all. The trained network operates at full resolution and runs in constant time. We investigate the effect of network architecture on approximation accuracy, runtime, and memory footprint, and identify a specific architecture that balances these considerations. We evaluate the presented approach on ten advanced image processing operators, including multiple variational models, multiscale tone and detail manipulation, photographic style transfer, nonlocal dehazing, and nonphotorealistic stylization. All operators are approximated by the same model. Experiments demonstrate that the presented approach is significantly more accurate than prior approximation schemes. It increases approximation accuracy as measured by PSNR across the evaluated operators by 8.5 dB on the MIT-Adobe dataset (from 27.5 to 36 dB) and reduces DSSIM by a multiplicative factor of 3 compared to the most accurate prior approximation scheme, while being the fastest. We show that our models generalize across datasets and across resolutions, and investigate a number of extensions of the presented approach.

We consider the problem of face swapping in images, where an input identity is transformed into a target identity while preserving pose, facial expression and lighting. To perform this mapping, we use convolutional neural networks trained to capture the appearance of the target identity from an unstructured collection of his/her photographs. This approach is enabled by framing the face swapping problem in terms of style transfer, where the goal is to render an image in the style of another one. Building on recent advances in this area, we devise a new loss function that enables the network to produce highly photorealistic results. By combining neural networks with simple pre- and post-processing steps, we aim at making face swap work in real-time with no input from the user.

What is the future of fashion? Tackling this question from a data-driven vision perspective, we propose to forecast visual style trends before they occur. We introduce the first approach to predict the future popularity of styles discovered from fashion images in an unsupervised manner. Using these styles as a basis, we train a forecasting model to represent their trends over time. The resulting model can hypothesize new mixtures of styles that will become popular in the future, discover style dynamics (trendy vs. classic), and name the key visual attributes that will dominate tomorrow's fashion. We demonstrate our idea applied to three datasets encapsulating 80,000 fashion products sold across six years on Amazon. Results indicate that fashion forecasting benefits greatly from visual analysis, much more than textual or meta-data cues surrounding products.

Although Deep Convolutional Neural Networks (CNNs) have liberated their power in various computer vision tasks, the most important components of CNN, convolutional layers and fully connected layers, are still limited to linear transformations. In this paper, we propose a novel Factorized Bilinear (FB) layer to model the pairwise feature interactions by considering the quadratic terms in the transformations. Compared with existing methods that tried to incorporate complex non-linearity structures into CNNs, the factorized parameterization makes our FB layer only require a linear increase of parameters and affordable computational cost. To further reduce the risk of overfitting of the FB layer, a specific remedy called DropFactor is devised during the training process. We also analyze the connection between FB layer and some existing models, and show FB layer is a generalization to them. Finally, we validate the effectiveness of FB layer on several widely adopted datasets including CIFAR-10, CIFAR-100 and ImageNet, and demonstrate superior results compared with various state-of-the-art deep models.

Face sketch to digital image matching is an important challenge of face recognition that involves matching across different domains. Current research efforts have primarily focused on extracting domain invariant representations or learning a mapping from one domain to the other. In this research, we propose a novel transform learning based approach termed as DeepTransformer, which learns a transformation and mapping function between the features of two domains. The proposed formulation is independent of the input information and can be applied with any existing learned or hand-crafted feature. Since the mapping function is directional in nature, we propose two variants of DeepTransformer: (i) semi-coupled and (ii) symmetrically-coupled deep transform learning. This research also uses a novel IIIT-D Composite Sketch with Age (CSA) variations database which contains sketch images of 150 subjects along with age-separated digital photos. The performance of the proposed models is evaluated on a novel application of sketch-to-sketch matching, along with sketch-to-digital photo matching. Experimental results demonstrate the robustness of the proposed models in comparison to existing state-of-the-art sketch matching algorithms and a commercial face recognition system.

We propose a lightweight method for dense online monocular depth estimation capable of reconstructing 3D meshes on computationally constrained platforms. Our main contribution is to pose the reconstruction problem as a non-local variational optimization over a time-varying Delaunay graph of the scene geometry, which allows for an efficient, keyframeless approach to depth estimation. The graph can be tuned to favor reconstruction quality or speed and is continuously smoothed and augmented as the camera explores the scene. Unlike keyframe-based approaches, the optimized surface is always available at the current pose, which is necessary for low-latency obstacle avoidance. FLaME (Fast Lightweight Mesh Estimation) can generate mesh reconstructions at upwards of 230 Hz using less than one Intel i7 CPU core, which enables operation on size, weight, and power-constrained platforms. We present results from both benchmark datasets and experiments running FLaME in-the-loop onboard a small flying quadrotor.

In this paper, we develop deep spatio-temporal neural networks to sequentially count vehicles from low quality videos captured by city cameras (citycams). Citycam videos have low resolution, low frame rate, high occlusion and large perspective, making most existing methods lose their efficacy. To overcome limitations of existing methods and incorporate the temporal information of traffic video, we design a novel FCN-rLSTM network to jointly estimate vehicle density and vehicle count by connecting fully convolutional neural networks (FCN) with long short term memory networks (LSTM) in a residual learning fashion. Such design leverages the strengths of FCN for pixel-level prediction and the strengths of LSTM for learning complex temporal dynamics. The residual learning connection reformulates the vehicle count regression as learning residual functions with reference to the sum of densities in each frame, which significantly accelerates the training of networks. To preserve feature map resolution, we propose a Hyper-Atrous combination to integrate atrous convolution in FCN and combine feature maps of different convolution layers. FCN-rLSTM enables refined feature representation and a novel end-to-end trainable mapping from pixels to vehicle count. We extensively evaluated the proposed method on different counting tasks with three datasets, with experimental results demonstrating their effectiveness and robustness. In particular, FCN-rLSTM reduces the mean absolute error (MAE) from 5.31 to 4.21 on TRANCOS; and reduces the MAE from 2.74 to 1.53 on WebCamT. Training process is accelerated by 5 times on average.

Manually annotating object bounding boxes is central to building computer vision datasets, and it is very time consuming (annotating ILSVRC [53] took 35s for one high-quality box [62]). It involves clicking on imaginary corners of a tight box around the object. This is difficult as these corners are often outside the actual object and several adjustments are required to obtain a tight box. We propose extreme clicking instead: we ask the annotator to click on four physical points on the object: the top, bottom, left- and right-most points. This task is more natural and these points are easy to find. We crowd-source extreme point annotations for PASCAL VOC 2007 and 2012 and show that (1) annotation time is only 7s per box, 5x faster than the traditional way of drawing boxes [62]; (2) the quality of the boxes is as good as the original ground-truth drawn the traditional way; (3) detectors trained on our annotations are as accurate as those trained on the original ground-truth. Moreover, our extreme clicking strategy not only yields box coordinates, but also four accurate boundary points. We show (4) how to incorporate them into GrabCut to obtain more accurate segmentations than those delivered when initializing it from bounding boxes; (5) semantic segmentations models trained on these segmentations outperform those trained on segmentations derived from bounding boxes.

The advent of high-tech journaling tools facilitates an image to be manipulated in a way that can easily evade state-of-the-art image tampering detection approaches. The recent success of the deep learning approaches in different recognition tasks inspires us to develop a high confidence detection framework which can localize manipulated regions in an image. Unlike semantic object segmentation where all meaningful regions (objects) are segmented, the localization of image manipulation focuses only the possible tampered region which makes the problem even more challenging. In order to formulate the framework, we employ a hybrid CNN-LSTM model to capture discriminative features between manipulated and non-manipulated regions. One of the key properties of manipulated regions is that they exhibit discriminative features in boundaries shared with neighboring non-manipulated pixels. Our motivation is to learn the boundary discrepancy, i.e., the spatial structure, between manipulated and non-manipulated regions with the combination of LSTM and convolution layers. We perform end-to-end training of the network to learn the parameters through back-propagation given groundtruth mask information. The overall framework is capable of detecting different types of image manipulations, including copy-move, removal and splicing. Our model shows promising results in localizing manipulated regions, which is demonstrated through rigorous experimentation on three diverse datasets.

Precise search of visually-similar vehicles poses a great challenge in computer vision, which needs to find exactly the same vehicle among a massive vehicles with visually similar appearances for a given query image. In this paper, we model the relationship of vehicle images as multiple grains. Following this, we propose two approaches to alleviate the precise vehicle search problem by exploiting multi-grain ranking constraints. One is Generalized Pairwise Ranking, which generalizes the conventional pairwise from considering only binary similar/dissimilar relations to multiple relations. The other is Multi-Grain based List Ranking, which introduces permutation probability to score a permutation of a multi-grain list, and further optimizes the ranking by the likelihood loss function. We implement the two approaches with multi-attribute classification in a multi-task deep learning framework. To further facilitate the research on precise vehicle search, we also contribute two high-quality and well-annotated vehicle datasets, named VD1 and VD2, which are collected from two different cities with diverse annotated attributes. As two of the largest publicly available precise vehicle search datasets, they contain 1,097,649 and 807,260 vehicle images respectively. Experimental results show that our approaches achieve the state-of-the-art performance on both datasets.