CVPR 2019 Papers

We investigate the problem of 6 degrees of freedom (DOF) camera planning for filming professional human motion videos using a camera drone. Existing methods either plan motions for only a pan-tilt-zoom (PTZ) camera, or adopt ad-hoc solutions without carefully considering the impact of video contents and previous camera motions on the future camera motions. As a result, they can hardly achieve satisfactory results in our drone cinematography task. In this study, we propose a learning-based framework which incorporates the video contents and previous camera motions to predict the future camera motions that enable the capture of professional videos. Specifically, the inputs of our framework are video contents which are represented using subject-related feature based on 2D skeleton and scene-related features extracted from background RGB images, and camera motions which are represented using optical flows. The correlation between the inputs and output future camera motions are learned via a sequence-to-sequence convolutional long short-term memory (Seq2Seq ConvLSTM) network from a large set of video clips. We deploy our approach to a real drone cinematography system by first predicting the future camera motions, and then converting them to the drone's control commands via an odometer. Our experimental results on extensive datasets and showcases exhibit significant improvements in our approach over conventional baselines and our approach can successfully mimic the footage of a professional cameraman.

In this paper, we introduce the task of generating a sharp slow-motion video given a low frame rate blurry video. We propose a data-driven approach, where the training data is captured with a high frame rate camera and blurry images are simulated through an averaging process. While it is possible to train a neural network to recover the sharp frames from their average, there is no guarantee of the temporal smoothness for the formed video, as the frames are estimated independently. To address the temporal smoothness requirement we propose a system with two networks: One, DeblurNet, to predict sharp keyframes and the second, InterpNet, to predict intermediate frames between the generated keyframes. A smooth transition is ensured by interpolating between consecutive keyframes using InterpNet. Moreover, the proposed scheme enables further increase in frame rate without retraining the network, by applying InterpNet recursively between pairs of sharp frames. We evaluate the proposed method on several datasets, including a novel dataset captured with a Sony RX V camera. We also demonstrate its performance of increasing the frame rate up to 20 times on real blurry videos.

The human visual system excels at biasing the stereoscopic visual signals by the attention mechanisms. Traditional methods relying on the low-level features and depth relevant information for stereoscopic video saliency prediction have fundamental limitations. For example, it is cumbersome to model the interactions between multiple visual cues including spatial, temporal, and depth information as a result of the sophistication. In this paper, we argue that the high-level features are crucial and resort to the deep learning framework to learn the saliency map of stereoscopic videos. Driven by spatio-temporal coherence from consecutive frames, the model first imitates the mechanism of saliency by taking advantage of the 3D convolutional neural network. Subsequently, the saliency originated from the intrinsic depth is derived based on the correlations between left and right views in a data-driven manner. Finally, a Convolutional Long Short-Term Memory (Conv-LSTM) based fusion network is developed to model the instantaneous interactions between spatio-temporal and depth attributes, such that the ultimate stereoscopic saliency maps over time are produced. Moreover, we establish a new large-scale stereoscopic video saliency dataset (SVS) including 175 stereoscopic video sequences and their fixation density annotations, aiming to comprehensively study the intrinsic attributes for stereoscopic video saliency detection. Extensive experiments show that our proposed model can achieve superior performance compared to the state-of-the-art methods on the newly built dataset for stereoscopic videos.

The recent advances in instance-level detection tasks lay a strong foundation for automated visual scenes understanding. However, the ability to fully comprehend a social scene still eludes us. In this work, we focus on detecting human-object interactions (HOIs) in images, an essential step towards deeper scene understanding. HOI detection aims to localize human and objects, as well as to identify the complex interactions between them. Innate in practical problems with large label space, HOI categories exhibit a long-tail distribution, i.e., there exist some rare categories with very few training samples. Given the key observation that HOIs contain intrinsic semantic regularities despite they are visually diverse, we tackle the challenge of long-tail HOI categories by modeling the underlying regularities among verbs and objects in HOIs as well as general relationships. In particular, we construct a knowledge graph based on the ground-truth annotations of training dataset and external source. In contrast to direct knowledge incorporation, we address the necessity of dynamic image-specific knowledge retrieval by multi-modal learning, which leads to an enhanced semantic embedding space for HOI comprehension. The proposed method shows improved performance on V-COCO and HICO-DET benchmarks, especially when predicting the rare HOI categories.

This paper presents a new deep-learning based method to simultaneously calibrate the intrinsic parameters of fisheye lens and rectify the distorted images. Assuming that the distorted lines generated by fisheye projection should be straight after rectification, we propose a novel deep neural network to impose explicit geometry constraints onto processes of the fisheye lens calibration and the distorted image rectification. In addition, considering the nonlinearity of distortion distribution in fisheye images, the proposed network fully exploits multi-scale perception to equalize the rectification effects on the whole image. To train and evaluate the proposed model, we also create a new large-scale dataset labeled with corresponding distortion parameters and well-annotated distorted lines. Compared with the state-of-the-art methods, our model achieves the best published rectification quality and the most accurate estimation of distortion parameters on a large set of synthetic and real fisheye images.

Real world applications of stereo depth estimation require models that are robust to dynamic variations in the environment. Even though deep learning based stereo methods are successful, they often fail to generalize to unseen variations in the environment, making them less suitable for practical applications such as autonomous driving. In this work, we introduce a "learning-to-adapt" framework that enables deep stereo methods to continuously adapt to new target domains in an unsupervised manner. Specifically, our approach incorporates the adaptation procedure into the learning objective to obtain a base set of parameters that are better suited for unsupervised online adaptation. To further improve the quality of the adaptation, we learn a confidence measure that effectively masks the errors introduced during the unsupervised adaptation. We evaluate our method on synthetic and real-world stereo datasets and our experiments evidence that learning-to-adapt is, indeed beneficial for online adaptation on vastly different domains.

We propose a generic framework to calibrate accuracy and confidence of a prediction in deep neural networks through stochastic inferences. We interpret stochastic regularization using a Bayesian model, and analyze the relation between predictive uncertainty of networks and variance of the prediction scores obtained by stochastic inferences for a single example. Our empirical study shows that the accuracy and the score of a prediction are highly correlated with the variance of multiple stochastic inferences given by stochastic depth or dropout. Motivated by this observation, we design a novel variance-weighted confidence-integrated loss function that is composed of two cross-entropy loss terms with respect to ground-truth and uniform distribution, which are balanced by variance of stochastic prediction scores. The proposed loss function enables us to learn deep neural networks that predict confidence calibrated scores using a single inference. Our algorithm presents outstanding confidence calibration performance and improves classification accuracy when combined with two popular stochastic regularization techniques---stochastic depth and dropout---in multiple models and datasets; it alleviates overconfidence issue in deep neural networks significantly by training networks to achieve prediction accuracy proportional to confidence of prediction.

Conventionally, deep neural networks are trained offline, relying on a large dataset prepared in advance. This paradigm is often challenged in real-world applications, e.g. online services that involve continuous streams of incoming data. Recently, incremental learning receives increasing attention, and is considered as a promising solution to the practical challenges mentioned above. However, it has been observed that incremental learning is subject to a fundamental difficulty -- catastrophic forgetting, namely adapting a model to new data often results in severe performance degradation on previous tasks or classes. Our study reveals that the imbalance between previous and new data is a crucial cause to this problem. In this work, we develop a new framework for incrementally learning a unified classifier, e.g. a classifier that treats both old and new classes uniformly. Specifically, we incorporate three components, cosine normalization, less-forget constraint, and inter-class separation, to mitigate the adverse effects of the imbalance. Experiments show that the proposed method can effectively rebalance the training process, thus obtaining superior performance compared to the existing methods. On CIFAR-100 and ImageNet, our method can reduce the classification errors by more than 6% and 13% respectively, under the incremental setting of 10 phases.

Deep ConvNets have shown great performance for single-label image classification (e.g. ImageNet), but it is necessary to move beyond the single-label classification task because pictures of everyday life are inherently multi-label. Multi-label classification is a more difficult task than single-label classification because both the input images and output label spaces are more complex. Furthermore, collecting clean multi-label annotations is more difficult to scale-up than single-label annotations. To reduce the annotation cost, we propose to train a model with partial labels i.e. only some labels are known per image. We first empirically compare different labeling strategies to show the potential for using partial labels on multi-label datasets. Then to learn with partial labels, we introduce a new classification loss that exploits the proportion of known labels per example. Our approach allows the use of the same training settings as when learning with all the annotations. We further explore several curriculum learning based strategies to predict missing labels. Experiments are performed on three large-scale multi-label datasets: MS COCO, NUS-WIDE and Open Images.

We propose a framework for learning through drawing. Our goal is to learn the correspondence between spoken words and abstract visual attributes, from a dataset of spoken descriptions of images. Building upon recent findings that GAN representations can be manipulated to edit semantic concepts in the generated output, we propose a new method to use such GAN-generated images to train a model using a triplet loss. To apply the method, we develop Audio CLEVRGAN, a new dataset of audio descriptions of GAN-generated CLEVR images, and we describe a training procedure that creates a curriculum of GAN-generated images that focuses training on image pairs that differ in a specific, informative way. Training is done without additional supervision beyond the spoken captions and the GAN. We find that training that takes advantage of GAN-generated edited examples results in improvements in the model's ability to learn attributes compared to previous results. Our proposed learning framework also results in models that can associate spoken words with some abstract visual concepts such as color and size.

Incremental learning (IL) is an important task aimed at increasing the capability of a trained model, in terms of the number of classes recognizable by the model. The key problem in this task is the requirement of storing data (e.g. images) associated with existing classes, while teaching the classifier to learn new classes. However, this is impractical as it increases the memory requirement at every incremental step, which makes it impossible to implement IL algorithms on edge devices with limited memory. Hence, we propose a novel approach, called `Learning without Memorizing (LwM)', to preserve the information about existing (base) classes, without storing any of their data, while making the classifier progressively learn the new classes. In LwM, we present an information preserving penalty: Attention Distillation Loss (L_ AD ), and demonstrate that penalizing the changes in classifiers' attention maps helps to retain information of the base classes, as new classes are added. We show that adding L_ AD to the distillation loss which is an existing information preserving loss consistently outperforms the state-of-the-art performance in the iILSVRC-small and iCIFAR-100 datasets in terms of the overall accuracy of base and incrementally learned classes.

Deep metric learning is essential for visual recognition. The widely used pair-wise (or triplet) based loss objectives cannot make full use of semantical information in training samples or give enough attention to those hard samples during optimization. Thus, they often suffer from a slow convergence rate and inferior performance. In this paper, we show how to learn an importance-driven distance metric via optimal transport programming from batches of samples. It can automatically emphasize hard examples and lead to significant improvements in convergence. We propose a new batch-wise optimal transport loss and combine it in an end-to-end deep metric learning manner. We use it to learn the distance metric and deep feature representation jointly for recognition. Empirical results on visual retrieval and classification tasks with six benchmark datasets, i.e., MNIST, CIFAR10, SHREC13, SHREC14, ModelNet10, and ModelNet40, demonstrate the superiority of the proposed method. It can accelerate the convergence rate significantly while achieving a state-of-the-art recognition performance. For example, in 3D shape recognition experiments, we show that our method can achieve better recognition performance within only 5 epochs than what can be obtained by mainstream 3D shape recognition approaches after 200 epochs.

There is some ambiguity in the 3D shape of an object when the number of observed views is small. Because of this ambiguity, although a 3D object reconstructor can be trained using a single view or a few views per object, reconstructed shapes only fit the observed views and appear incorrect from the unobserved viewpoints. To reconstruct shapes that look reasonable from any viewpoint, we propose to train a discriminator that learns prior knowledge regarding possible views. The discriminator is trained to distinguish the reconstructed views of the observed viewpoints from those of the unobserved viewpoints. The reconstructor is trained to correct unobserved views by fooling the discriminator. Our method outperforms current state-of-the-art methods on both synthetic and natural image datasets; this validates the effectiveness of our method.

This paper conducts a systematic study on the role of visual attention in Unsupervised Video Object Segmentation (UVOS) tasks. By elaborately annotating three popular video segmentation datasets (DAVIS, Youtube-Objects and SegTrack V2) with dynamic eye-tracking data in the UVOS setting, for the first time, we quantitatively verified the high consistency of visual attention behavior among human observers, and found strong correlation between human attention and explicit primary object judgements during dynamic, task-driven viewing. Such novel observations provide an in-depth insight into the underlying rationale behind UVOS. Inspired by these findings, we decouple UVOS into two sub-tasks: UVOS-driven Dynamic Visual Attention Prediction (DVAP) in spatiotemporal domain, and Attention-Guided Object Segmentation (AGOS) in spatial domain. Our UVOS solution enjoys three major merits: 1) modular training without using expensive video segmentation annotations, instead, using more affordable dynamic fixation data to train the initial video attention module and using existing fixation-segmentation paired static/image data to train the subsequent segmentation module; 2) comprehensive foreground understanding through multi-source learning; and 3) additional interpretability from the biologically-inspired and assessable attention. Experiments on popular benchmarks show that, even without using expensive video object mask annotations, our model achieves compelling performance in comparison with state-of-the-arts.

Reconstructing the 3D model of a physical object typically requires us to align the depth scans obtained from different camera poses into the same coordinate system. Solutions to this global alignment problem usually proceed in two steps. The first step estimates relative transformations between pairs of scans using an off-the-shelf technique. Due to limited information presented between pairs of scans, the resulting relative transformations are generally noisy. The second step then jointly optimizes the relative transformations among all input depth scans. A natural constraint used in this step is the cycle-consistency constraint, which allows us to prune incorrect relative transformations by detecting inconsistent cycles. The performance of such approaches, however, heavily relies on the quality of the input relative transformations. Instead of merely using the relative transformations as the input to perform transformation synchronization, we propose to use a neural network to learn the weights associated with each relative transformation. Our approach alternates between transformation synchronization using weighted relative transformations and predicting new weights of the input relative transformations using a neural network. We demonstrate the usefulness of this approach across a wide range of datasets.

Convolutional Neural Networks (CNN) have been regarded as a powerful class of models for visual recognition problems. Nevertheless, the convolutional filters in these networks are local operations while ignoring the large-range dependency. Such drawback becomes even worse particularly for video recognition, since video is an information-intensive media with complex temporal variations. In this paper, we present a novel framework to boost the spatio-temporal representation learning by Local and Global Diffusion (LGD). Specifically, we construct a novel neural network architecture that learns the local and global representations in parallel. The architecture is composed of LGD blocks, where each block updates local and global features by modeling the diffusions between these two representations. Diffusions effectively interact two aspects of information, i.e., localized and holistic, for more powerful way of representation learning. Furthermore, a kernelized classifier is introduced to combine the representations from two aspects for video recognition. Our LGD networks achieve clear improvements on the large-scale Kinetics-400 and Kinetics-600 video classification datasets against the best competitors by 3.5% and 0.7%. We further examine the generalization of both the global and local representations produced by our pre-trained LGD networks on four different benchmarks for video action recognition and spatio-temporal action detection tasks. Superior performances over several state-of-the-art techniques on these benchmarks are reported.

Depth estimation from a single image in the wild remains a challenging problem. One main obstacle is the lack of high-quality training data for images in the wild. In this paper we propose a method to automatically generate such data through Structure-from-Motion (SfM) on Internet videos. The core of this method is a Quality Assessment Network that identifies high-quality reconstructions obtained from SfM. Using this method, we collect single-view depth training data from a large number of YouTube videos and construct a new dataset called YouTube3D. Experiments show that YouTube3D is useful in training depth estimation networks and advances the state of the art of single-view depth estimation in the wild.

We address the problem of detecting scene text in arbitrary shapes, which is a challenging task due to the high variety and complexity of the scene. Specifically, we treat text detection as instance segmentation and propose a segmentation-based framework, which extracts each text instance as an independent connected component. To distinguish different text instances, our method maps pixels onto an embedding space where pixels belonging to the same text are encouraged to appear closer to each other and vise versa. In addition, we introduce a Shape-Aware Loss to make training adaptively accommodate various aspect ratios of text instances and the tiny gaps among them, and a new post-processing pipeline to yield precise bounding box predictions. Experimental results on three challenging datasets (ICDAR15, MSRA-TD500 and CTW1500) demonstrate the effectiveness of our work.

As an alternative to manual pixel-wise annotation, synthetic data has been increasingly used for training semantic segmentation models. Such synthetic images and semantic labels can be easily generated from virtual 3D environments. In this work, we propose an approach to cross-domain semantic segmentation with the auxiliary geometric information, which can also be easily obtained from virtual environments. The geometric information is utilized on two levels for reducing domain shift: on the input level, we augment the standard image translation network with the geometric information to translate synthetic images into realistic style; on the output level, we build a task network which simultaneously performs semantic segmentation and depth estimation. Meanwhile, adversarial training is applied on the joint output space to preserve the correlation between semantics and depth. The proposed approach is validated on two pairs of synthetic to real dataset: from Virtual KITTI to KITTI, and from SYNTHIA to Cityscapes, where we achieve a clear performance gain compared to the baselines and various competing methods, demonstrating the effectiveness of the geometric information for cross-domain semantic segmentation.