ICCV 2019 Papers

Adversarial robustness has emerged as an important topic in deep learning as carefully crafted attack samples can significantly disturb the performance of a model. Many recent methods have proposed to improve adversarial robustness by utilizing adversarial training or model distillation, which adds additional procedures to model training. In this paper, we propose a new training paradigm called Guided Complement Entropy (GCE) that is capable of achieving "adversarial defense for free," which involves no additional procedures in the process of improving adversarial robustness. In addition to maximizing model probabilities on the ground-truth class like cross-entropy, we neutralize its probabilities on the incorrect classes along with a "guided" term to balance between these two terms. We show in the experiments that our method achieves better model robustness with even better performance compared to the commonly used cross-entropy training objective. We also show that our method can be used orthogonal to adversarial training across well-known methods with noticeable robustness gain. To the best of our knowledge, our approach is the first one that improves model robustness without compromising performance.

Adaptive inference is a promising technique to improve the computational efficiency of deep models at test time. In contrast to static models which use the same computation graph for all instances, adaptive networks can dynamically adjust their structure conditioned on each input. While existing research on adaptive inference mainly focuses on designing more advanced architectures, this paper investigates how to train such networks more effectively. Specifically, we consider a typical adaptive deep network with multiple intermediate classifiers. We present three techniques to improve its training efficacy from two aspects: 1) a Gradient Equilibrium algorithm to resolve the conflict of learning of different classifiers; 2) an Inline Subnetwork Collaboration approach and a One-for-all Knowledge Distillation algorithm to enhance the collaboration among classifiers. On multiple datasets (CIFAR-10, CIFAR-100 and ImageNet), we show that the proposed approach consistently leads to further improved efficiency on top of state-of-the-art adaptive deep networks.

Predicting future frames for a video sequence is a challenging generative modeling task. Promising approaches include probabilistic latent variable models such as the Variational Auto-Encoder. While VAEs can handle uncertainty and model multiple possible future outcomes, they have a tendency to produce blurry predictions. In this work we argue that this is a sign of underfitting. To address this issue, we propose to increase the expressiveness of the latent distributions and to use higher capacity likelihood models. Our approach relies on a hierarchy of latent variables, which defines a family of flexible prior and posterior distributions in order to better model the probability of future sequences. We validate our proposal through a series of ablation experiments and compare our approach to current state-of-the-art latent variable models. Our method performs favorably under several metrics in three different datasets.

Implicit shape representations, such as Level Sets, provide a very elegant formulation for performing computations involving curves and surfaces. However, including implicit representations into canonical Neural Network formulations is far from straightforward. This has consequently restricted existing approaches to shape inference, to significantly less effective representations, perhaps most commonly voxels occupancy maps or sparse point clouds. To overcome this limitation we propose a novel formulation that permits the use of implicit representations of curves and surfaces, of arbitrary topology, as individual layers in Neural Network architectures with end-to-end trainability. Specifically, we propose to represent the output as an oriented level set of a continuous and discretised embedding function. We investigate the benefits of our approach on the task of 3D shape prediction from a single image; and demonstrate its ability to produce a more accurate reconstruction compared to voxel-based representations. We further show that our model is flexible and can be applied to a variety of shape inference problems.

Modeling and prediction of human motion dynamics has long been a challenging problem in computer vision, and most existing methods rely on the end-to-end supervised training of various architectures of recurrent neural networks. Inspired by the recent success of deep reinforcement learning methods, in this paper we propose a new reinforcement learning formulation for the problem of human pose prediction, and develop an imitation learning algorithm for predicting future poses under this formulation through a combination of behavioral cloning and generative adversarial imitation learning. Our experiments show that our proposed method outperforms all existing state-of-the-art baseline models by large margins on the task of human pose prediction in both short-term predictions and long-term predictions, while also enjoying huge advantage in training speed.

Despite remarkable recent progress on both unconditional and conditional image synthesis, it remains a long- standing problem to learn generative models that are capable of synthesizing realistic and sharp images from re- configurable spatial layout (i.e., bounding boxes + class labels in an image lattice) and style (i.e., structural and appearance variations encoded by latent vectors), especially at high resolution. By reconfigurable, it means that a model can preserve the intrinsic one-to-many mapping from a given layout to multiple plausible images with different styles, and is adaptive with respect to perturbations of a layout and style latent code. In this paper, we present a layout- and style-based architecture for generative adversarial networks (termed LostGANs) that can be trained end-to-end to generate images from reconfigurable layout and style. Inspired by the vanilla StyleGAN, the proposed LostGAN consists of two new components: (i) learning fine-grained mask maps in a weakly-supervised manner to bridge the gap between layouts and images, and (ii) learning object instance-specific layout-aware feature normalization (ISLA-Norm) in the generator to realize multi-object style generation. In experiments, the proposed method is tested on the COCO-Stuff dataset and the Visual Genome dataset with state-of-the-art performance obtained. The code and pretrained models are available at https://github.com/iVMCL/LostGANs.

Most convolutional network (CNN)-based inpainting methods adopt standard convolution to indistinguishably treat valid pixels and holes, making them limited in handling irregular holes and more likely to generate inpainting results with color discrepancy and blurriness. Partial convolution has been suggested to address this issue, but it adopts handcrafted feature re-normalization, and only considers forward mask-updating. In this paper, we present a learnable attention map module for learning feature re-normalization and mask-updating in an end-to-end manner, which is effective in adapting to irregular holes and propagation of convolution layers. Furthermore, learnable reverse attention maps are introduced to allow the decoder of U-Net to concentrate on filling in irregular holes instead of reconstructing both holes and known regions, resulting in our learnable bidirectional attention maps. Qualitative and quantitative experiments show that our method performs favorably against state-of-the-arts in generating sharper, more coherent and visually plausible inpainting results. The source code and pre-trained models will be available at: https://github.com/Vious/LBAM_inpainting/.

Thanks to the recent development of deep generative models, it is becoming possible to generate high-quality images with both fidelity and diversity. However, the training of such generative models requires a large dataset. To reduce the amount of data required, we propose a new method for transferring prior knowledge of the pre-trained generator, which is trained with a large dataset, to a small dataset in a different domain. Using such prior knowledge, the model can generate images leveraging some common sense that cannot be acquired from a small dataset. In this work, we propose a novel method focusing on the parameters for batch statistics, scale and shift, of the hidden layers in the generator. By training only these parameters in a supervised manner, we achieved stable training of the generator, and our method can generate higher quality images compared to previous methods without collapsing, even when the dataset is small ( 100). Our results show that the diversity of the filters acquired in the pre-trained generator is important for the performance on the target domain. Our method makes it possible to add a new class or domain to a pre-trained generator without disturbing the performance on the original domain. Code is available at github.com/nogu-atsu/small-dataset-image-generation

We propose a unified approach to three tasks of aesthetic score regression, binary aesthetic classification, and personalized aesthetics. First, we develop a comparator to estimate the ratio of aesthetic scores for two images. Then, we construct a pairwise comparison matrix for multiple reference images and an input image, and predict the aesthetic score of the input via the eigenvalue decomposition of the matrix. By varying the reference images, the proposed algorithm can be used for binary aesthetic classification and personalized aesthetics, as well as generic score regression. Experimental results demonstrate that the proposed unified algorithm provides the state-of-the-art performances in all three tasks of image aesthetics.

We propose a novel approach that disentangles the identity and pose of objects for image generation. Our model takes as input an ID image and a pose image, and generates an output image with the identity of the ID image and the pose of the pose image. Unlike most previous unsupervised work which rely on cyclic constraints, which can often be brittle, we instead propose to learn this in a self-supervised way. Specifically, we leverage unlabeled videos to automatically construct pseudo ground-truth targets to directly supervise our model. To enforce disentanglement, we propose a novel disentanglement loss, and to improve realism, we propose a pixel-verification loss in which the generated image's pixels must trace back to the ID input. We conduct extensive experiments on both synthetic and real images to demonstrate improved realism, diversity, and ID/pose disentanglement compared to existing methods.

In this work, we present a new operator, called Instance Mask Projection (IMP), which projects a predicted instance segmentation as a new feature for semantic segmentation. It also supports back propagation and is trainable end-to end. By adding this operator, we introduce a new way to combine top-down and bottom-up information in semantic segmentation. Our experiments show the effectiveness of IMP on both clothing parsing (with complex layering, large deformations, and non-convex objects), and on street scene segmentation (with many overlapping instances and small objects). On the Varied Clothing Parsing dataset (VCP), we show instance mask projection can improve mIOU by 3 points over a state-of-the-art Panoptic FPN segmentation approach. On the ModaNet clothing parsing dataset, we show a dramatic improvement of 20.4% compared to existing baseline semantic segmentation results. In addition, the Instance Mask Projection operator works well on other (non-clothing) datasets, providing an improvement in mIOU of 3 points on "thing" classes of Cityscapes, a self-driving dataset, over a state-of-the-art approach.

This paper presents a class incremental learning (IL) method which exploits fine tuning and a dual memory to reduce the negative effect of catastrophic forgetting in image recognition. First, we simplify the current fine tuning based approaches which use a combination of classification and distillation losses to compensate for the limited availability of past data. We find that the distillation term actually hurts performance when a memory is allowed. Then, we modify the usual class IL memory component. Similar to existing works, a first memory stores exemplar images of past classes. A second memory is introduced here to store past class statistics obtained when they were initially learned. The intuition here is that classes are best modeled when all their data are available and that their initial statistics are useful across different incremental states. A prediction bias towards newly learned classes appears during inference because the dataset is imbalanced in their favor. The challenge is to make predictions of new and past classes more comparable. To do this, scores of past classes are rectified by leveraging contents from both memories. The method has negligible added cost, both in terms of memory and of inference complexity. Experiments with three large public datasets show that the proposed approach is more effective than a range of competitive state-of-the-art methods.

To solve the low spatial and/or temporal resolution problem which the conventional hypelrspectral cameras often suffer from, coded snapshot hyperspectral imaging systems have attracted more attention recently. Recovering a hyperspectral image (HSI) from its corresponding coded image is an ill-posed inverse problem, and learning accurate prior of HSI is essential to solve this inverse problem. In this paper, we present an effective convolutional neural network (CNN) based method for coded HSI reconstruction, which learns the deep prior from the external dataset as well as the internal information of input coded image with spatial-spectral constraint. Our method can effectively exploit spatial-spectral correlation and sufficiently represent the variety nature of HSIs. Experimental results show our method outperforms the state-of-the-art methods under both comprehensive quantitative metrics and perceptive quality.

Establishing visual correspondences under large intra-class variations requires analyzing images at different levels, from features linked to semantics and context to local patterns, while being invariant to instance-specific details. To tackle these challenges, we represent images by "hyperpixels" that leverage a small number of relevant features selected among early to late layers of a convolutional neural network. Taking advantage of the condensed features of hyperpixels, we develop an effective real-time matching algorithm based on Hough geometric voting. The proposed method, hyperpixel flow, sets a new state of the art on three standard benchmarks as well as a new dataset, SPair-71k, which contains a significantly larger number of image pairs than existing datasets, with more accurate and richer annotations for in-depth analysis.

This paper proposes a human-aware deblurring model that disentangles the motion blur between foreground (FG) humans and background (BG). The proposed model is based on a triple-branch encoder-decoder architecture. The first two branches are learned for sharpening FG humans and BG details, respectively; while the third one produces global, harmonious results by comprehensively fusing multi-scale deblurring information from the two domains. The proposed model is further endowed with a supervised, human-aware attention mechanism in an end-to-end fashion. It learns a soft mask that encodes FG human information and explicitly drives the FG/BG decoder-branches to focus on their specific domains. Above designs lead to a fully differentiable motion deblurring network, which can be trained end-to-end. To further benefit the research towards Human-aware Image Deblurring, we introduce a large-scale dataset, named HIDE, which consists of 8,422 blurry and sharp image pairs with 65,784 densely annotated FG human bounding boxes. HIDE is specifically built to span a broad range of scenes, human object sizes, motion patterns, and background complexities. Extensive experiments on public benchmarks and our dataset demonstrate that our model performs favorably against the state-of-the-art motion deblurring methods, especially in capturing semantic details.

The classification performance of deep neural networks has begun to asymptote at near-perfect levels. However, their ability to generalize outside the training set and their robustness to adversarial attacks have not. In this paper, we make progress on this problem by training with full label distributions that reflect human perceptual uncertainty. We first present a new benchmark dataset which we call CIFAR10H, containing a full distribution of human labels for each image of the CIFAR10 test set. We then show that, while contemporary classifiers fail to exhibit human-like uncertainty on their own, explicit training on our dataset closes this gap, supports improved generalization to increasingly out-of-training-distribution test datasets, and confers robustness to adversarial attacks.

We propose a Generative Adversarial Network (GAN) to forecast 3D human motion given a sequence of past 3D skeleton poses. While recent GANs have shown promising results, they can only forecast plausible motion over relatively short periods of time (few hundred milliseconds) and typically ignore the absolute position of the skeleton w.r.t. the camera. Our scheme provides long term predictions (two seconds or more) for both the body pose and its absolute position. Our approach builds upon three main contributions. First, we represent the data using a spatio-temporal tensor of 3D skeleton coordinates which allows formulating the prediction problem as an inpainting one, for which GANs work particularly well. Secondly, we design an architecture to learn the joint distribution of body poses and global motion, capable to hypothesize large chunks of the input 3D tensor with missing data. And finally, we argue that the L2 metric, considered so far by most approaches, fails to capture the actual distribution of long-term human motion. We propose two alternative metrics, based on the distribution of frequencies, that are able to capture more realistic motion patterns. Extensive experiments demonstrate our approach to significantly improve the state of the art, while also handling situations in which past observations are corrupted by occlusions, noise and missing frames.

We describe an end-to-end method for recovering 3D human body mesh from single images and monocular videos. Different from the existing methods try to obtain all the complex 3D pose, shape, and camera parameters from one coupling feature, we propose a skeleton-disentangling based framework, which divides this task into multi-level spatial and temporal granularity in a decoupling manner. In spatial, we propose an effective and pluggable "disentangling the skeleton from the details" (DSD) module. It reduces the complexity and decouples the skeleton, which lays a good foundation for temporal modeling. In temporal, the self-attention based temporal convolution network is proposed to efficiently exploit the short and long-term temporal cues. Furthermore, an unsupervised adversarial training strategy, temporal shuffles and order recovery, is designed to promote the learning of motion dynamics. The proposed method outperforms the state-of-the-art 3D human mesh recovery methods by 15.4% MPJPE and 23.8% PA-MPJPE on Human3.6M. State-of-the-art results are also achieved on the 3D pose in the wild (3DPW) dataset without any fine-tuning. Especially, ablation studies demonstrate that skeleton-disentangled representation is crucial for better temporal modeling and generalization.

In this work, we present a novel dataset consisting of eye movements and verbal descriptions recorded synchronously over images. Using this data, we study the differences in human attention during free-viewing and image captioning tasks. We look into the relationship between human atten- tion and language constructs during perception and sen- tence articulation. We also analyse attention deployment mechanisms in the top-down soft attention approach that is argued to mimic human attention in captioning tasks, and investigate whether visual saliency can help image caption- ing. Our study reveals that (1) human attention behaviour differs in free-viewing and image description tasks. Hu- mans tend to fixate on a greater variety of regions under the latter task, (2) there is a strong relationship between de- scribed objects and attended objects (97% of the described objects are being attended), (3) a convolutional neural net- work as feature encoder accounts for human-attended re- gions during image captioning to a great extent (around 78%), (4) soft-attention mechanism differs from human at- tention, both spatially and temporally, and there is low correlation between caption scores and attention consis- tency scores. These indicate a large gap between humans and machines in regards to top-down attention, and (5) by integrating the soft attention model with image saliency, we can significantly improve the model's performance on Flickr30k and MSCOCO benchmarks. The dataset can be found at: https://github.com/SenHe/ Human-Attention-in-Image-Captioning.