CVPR 2020 Papers

In the field of face recognition, large-scale web-collected datasets are essential for learning discriminative representations, but they suffer from noisy identity labels, such as outliers and label flips. It is beneficial to automatically cleanse their label noise for improving recognition accuracy. Unfortunately, existing cleansing methods cannot accurately identify noise in the wild. To solve this problem, we propose an effective automatic label noise cleansing framework for face recognition datasets, FaceGraph. Using two cascaded graph convolutional networks, FaceGraph performs global-to-local discrimination to select useful data in a noisy environment. Extensive experiments show that cleansing widely used datasets, such as CASIA-WebFace, VGGFace2, MegaFace2, and MS-Celeb-1M, using the proposed method can improve the recognition performance of state-of-the-art representation learning methods like Arcface. Further, we cleanse massive self-collected celebrity data, namely MillionCelebs, to provide 18.8M images of 636K identities. Training with the new data, Arcface surpasses state-of-the-art performance by a notable margin to reach 95.62% TPR at 1e-5 FPR on the IJB-C benchmark.

Local and global patterns of an object are closely related. Although each part of an object is incomplete, the underlying attributes about the object are shared among all parts, which makes reasoning the whole object from a single part possible. We hypothesize that a powerful representation of a 3D object should model the attributes that are shared between parts and the whole object, and distinguishable from other objects. Based on this hypothesis, we propose to learn point cloud representation by bidirectional reasoning between the local structures at different abstraction hierarchies and the global shape without human supervision. Experimental results on various benchmark datasets demonstrate the unsupervisedly learned representation is even better than supervised representation in discriminative power, generalization ability, and robustness. We show that unsupervisedly trained point cloud models can outperform their supervised counterparts on downstream classification tasks. Most notably, by simply increasing the channel width of an SSG PointNet++, our unsupervised model surpasses the state-of-the-art supervised methods on both synthetic and real-world 3D object classification datasets. We expect our observations to offer a new perspective on learning better representation from data structures instead of human annotations for point cloud understanding.

Generative Adversarial Networks (GANs) can generate realistic fake face images that can easily fool human beings. On the contrary, a common Convolutional Neural Network(CNN) discriminator can achieve more than99.9%accuracyin discerning fake/real images. In this paper, we conduct an empirical study on fake/real faces, and have two important observations: firstly, the texture of fake faces is substantially different from real ones; secondly, global texture statistics are more robust to image editing and transferable to fake faces from different GANs and datasets. Motivated by the above observations, we propose a new architecture coined as Gram-Net, which leverages global image texture representations for robust fake image detection. Experimental results on several datasets demonstrate that our Gram-Netoutperforms existing approaches. Especially, our Gram-Netis more robust to image editings, e.g. down-sampling, JPEGcompression, blur, and noise. More importantly, our Gram-Net generalizes significantly better in detecting fake faces from GAN models not seen in the training phase and can perform decently in detecting fake natural images

In this paper we present a study of global optimality conditions for Point Set Registration (PSR) with missing data. PSR is the problem of aligning multiple point clouds with an unknown target point cloud. Since non-linear rotation constraints are present the problem is inherently non-convex and typically relaxed by computing the Lagrange dual, which is a Semidefinite Program (SDP). In this work we show that given a local minimizer the dual variables of the SDP can be computed in closed form. This opens up the possibility of verifying the optimally, using the SDP formulation without explicitly solving it. In addition it allows us to study under what conditions the relaxation is tight, through spectral analysis. We show that if the errors in the (unknown) optimal solution are bounded the SDP formulation will be able to recover it.

Deploying convolutional neural networks (CNNs) on embedded devices is difficult due to the limited memory and computation resources. The redundancy in feature maps is an important characteristic of those successful CNNs, but has rarely been investigated in neural architecture design. This paper proposes a novel Ghost module to generate more feature maps from cheap operations. Based on a set of intrinsic feature maps, we apply a series of linear transformations with cheap cost to generate many ghost feature maps that could fully reveal information underlying intrinsic features. The proposed Ghost module can be taken as a plug-and-play component to upgrade existing convolutional neural networks. Ghost bottlenecks are designed to stack Ghost modules, and then the lightweight GhostNet can be easily established. Experiments conducted on benchmarks demonstrate that the proposed Ghost module is an impressive alternative of convolution layers in baseline models, and our GhostNet can achieve higher recognition performance (e.g. 75.7% top-1 accuracy) than MobileNetV3 with similar computational cost on the ImageNet ILSVRC-2012 classification dataset. Code is available at https://github.com/huawei-noah/ghostnet.

We present a novel method for generating panoramic street-view images which are geometrically consistent with a given satellite image. Different from existing approaches that completely rely on a deep learning architecture to generalize cross-view image distributions, our approach explicitly loops in the geometric configuration of the ground objects based on the satellite views, such that the produced ground view synthesis preserves the geometric shape and the semantics of the scene. In particular, we propose a neural network with a geo-transformation layer that turns predicted ground-height values from the satellite view to a ground view while retaining the physical satellite-to-ground relation. Our results show that the synthesized image retains well-articulated and authentic geometric shapes, as well as texture richness of the street-view in various scenarios. Both qualitative and quantitative results demonstrate that our method compares favorably to other state-of-the-art approaches that lack geometric consistency.

In this paper, we propose a normal estimation method for unstructured point cloud. We observe that geometric estimators commonly focus more on feature preservation but are hard to tune parameters and sensitive to noise, while learning-based approaches pursue an overall normal estimation accuracy but cannot well handle challenging regions such as surface edges. This paper presents a novel normal estimation method, under the co-support of geometric estimator and deep learning. To lowering the learning difficulty, we first propose to compute a suboptimal initial normal at each point by searching for a best fitting patch. Based on the computed normal field, we design a normal-based height map network (NH-Net) to fine-tune the suboptimal normals. Qualitative and quantitative evaluations demonstrate the clear improvements of our results over both traditional methods and learning-based methods, in terms of estimation accuracy and feature recovery.

Graph convolution operators bring the advantages of deep learning to a variety of graph and mesh processing tasks previously deemed out of reach. With their continued success comes the desire to design more powerful architectures, often by adapting existing deep learning techniques to non-Euclidean data. In this paper, we argue geometry should remain the primary driving force behind innovation in the emerging field of geometric deep learning. We relate graph neural networks to widely successful computer graphics and data approximation models: radial basis functions (RBFs). We conjecture that, like RBFs, graph convolution layers would benefit from the addition of simple functions to the powerful convolution kernels. We introduce affine skip connections, a novel building block formed by combining a fully connected layer with any graph convolution operator. We experimentally demonstrate the effectiveness of our technique, and show the improved performance is the consequence of more than the increased number of parameters. Operators equipped with the affine skip connection markedly outperform their base performance on every task we evaluated, i.e., shape reconstruction, dense shape correspondence, and graph classification. We hope our simple and effective approach will serve as a solid baseline and help ease future research in graph neural networks.

Motivated by the correlation between the depth and the geometric structure of a 360 indoor image, we propose a novel learning-based depth estimation framework that leverages the geometric structure of a scene to conduct depth estimation. Specifically, we represent the geometric structure of an indoor scene as a collection of corners, boundaries and planes. On the one hand, once a depth map is estimated, this geometric structure can be inferred from the estimated depth map; thus, the geometric structure functions as a regularizer for depth estimation. On the other hand, this estimation also benefits from the geometric structure of a scene estimated from an image where the structure functions as a prior. However, furniture in indoor scenes makes it challenging to infer geometric structure from depth or image data. An attention map is inferred to facilitate both depth estimation from features of the geometric structure and also geometric inferences from the estimated depth map. To validate the effectiveness of each component in our framework under controlled conditions, we render a synthetic dataset, Shanghaitech-Kujiale Indoor 360 dataset with 3550 360 indoor images. Extensive experiments on popular datasets validate the effectiveness of our solution. We also demonstrate that our method can also be applied to counterfactual depth.

Adversarial examples are known as carefully perturbed images fooling image classifiers. We propose a geometric framework to generate adversarial examples in one of the most challenging black-box settings where the adversary can only generate a small number of queries, each of them returning the top-1 label of the classifier. Our framework is based on the observation that the decision boundary of deep networks usually has a small mean curvature in the vicinity of data samples. We propose an effective iterative algorithm to generate query-efficient black-box perturbations with small p norms which is confirmed via experimental evaluations on state-of-the-art natural image classifiers. Moreover, for p=2, we theoretically show that our algorithm actually converges to the minimal perturbation when the curvature of the decision boundary is bounded. We also obtain the optimal distribution of the queries over the iterations of the algorithm. Finally, experimental results confirm that our principled black-box attack algorithm performs better than state-of-the-art algorithms as it generates smaller perturbations with a reduced number of queries.

We address the problem of open-set recognition, where the goal is to determine if a given sample belongs to one of the classes used for training a model (known classes). The main challenge in open-set recognition is to disentangle open-set samples that produce high class activations from known-set samples. We propose two techniques to force class activations of open-set samples to be low. First, we train a generative model for all known classes and then augment the input with the representation obtained from the generative model to learn a classifier. This network learns to associate high classification probabilities both when image content is from the correct class as well as when the input and the reconstructed image are consistent with each other. Second, we use self-supervision to force the network to learn more informative featues when assigning class scores to improve separation of classes from each other and from open-set samples. We evaluate the performance of the proposed method with recent open-set recognition works across three datasets, where we obtain state-of-the-art results.

Automatically reasoning about future human behaviors is a difficult problem but has significant practical applications to assistive systems. Part of this difficulty stems from learning systems' inability to represent all kinds of behaviors. Some behaviors, such as motion, are best described with continuous representations, whereas others, such as picking up a cup, are best described with discrete representations. Furthermore, human behavior is generally not fixed: people can change their habits and routines. This suggests these systems must be able to learn and adapt continuously. In this work, we develop an efficient deep generative model to jointly forecast a person's future discrete actions and continuous motions. On a large-scale egocentric dataset, EPIC-KITCHENS, we observe our method generates high-quality and diverse samples while exhibiting better generalization than related generative models. Finally, we propose a variant to continually learn our model from streaming data, observe its practical effectiveness, and theoretically justify its learning efficiency.

Beyond depth estimation from a single image, the monocular cue is useful in a broader range of depth inference applications and settings---such as when one can leverage other available depth cues for improved accuracy. Currently, different applications, with different inference tasks and combinations of depth cues, are solved via different specialized networks---trained separately for each application. Instead, we propose a versatile task-agnostic monocular model that outputs a probability distribution over scene depth given an input color image, as a sample approximation of outputs from a patch-wise conditional VAE. We show that this distributional output can be used to enable a variety of inference tasks in different settings, without needing to retrain for each application. Across a diverse set of applications (depth completion, user guided estimation, etc.), our common model yields results with high accuracy---comparable to or surpassing that of state-of-the-art methods dependent on application-specific networks.

Rectified Linear Units (ReLUs) are among the most widely used activation function in a broad variety of tasks in vision. Recent theoretical results suggest that despite their excellent practical performance, in various cases, a substitution with basis expansions (e.g., polynomials) can yield significant benefits from both the optimization and generalization perspective. Unfortunately, the existing results remain limited to networks with a couple of layers, and the practical viability of these results is not yet known. Motivated by some of these results, we explore the use of Hermite polynomial expansions as a substitute for ReLUs in deep networks. While our experiments with supervised learning do not provide a clear verdict, we find that this strategy offers considerable benefits in semi-supervised learning (SSL) / transductive learning settings. We carefully develop this idea and show how the use of Hermite polynomials based activations can yield improvements in pseudo-label accuracies and sizable financial savings (due to concurrent runtime benefits). Further, we show via theoretical analysis, that the networks (with Hermite activations) offer robustness to noise and other attractive mathematical properties.

We present a fully automatic system that takes a 3D scene and generates plausible 3D human bodies that are posed naturally in that 3D scene. Given a 3D scene without people, humans can easily imagine how people could interact with the scene and the objects in it. However, this is a challenging task for a computer as solving it requires that (1) the generated human bodies to be semantically plausible within the 3D environment (e.g. people sitting on the sofa or cooking near the stove), and (2) the generated human-scene interaction to be physically feasible such that the human body and scene do not interpenetrate while, at the same time, body-scene contact supports physical interactions. To that end, we make use of the surface-based 3D human model SMPL-X. We first train a conditional variational autoencoder to predict semantically plausible 3D human poses conditioned on latent scene representations, then we further refine the generated 3D bodies using scene constraints to enforce feasible physical interaction. We show that our approach is able to synthesize realistic and expressive 3D human bodies that naturally interact with 3D environment. We perform extensive experiments demonstrating that our generative framework compares favorably with existing methods, both qualitatively and quantitatively. We believe that our scene-conditioned 3D human generation pipeline will be useful for numerous applications; e.g. to generate training data for human pose estimation, in video games and in VR/AR. Our project page for data and code can be seen at: https://vlg.inf.ethz.ch/projects/PSI/ .

Although the performance of hand segmentation in egocentric videos has been significantly improved by using CNNs, it still remains a challenging issue to generalize the trained models to new domains, e.g., unseen environments. In this work, we solve the hand segmentation generalization problem without requiring segmentation labels in the target domain. To this end, we propose a Bayesian CNN-based model adaptation framework for hand segmentation, which introduces and considers two key factors: 1) prediction uncertainty when the model is applied in a new domain and 2) common information about hand shapes shared across domains. Consequently, we propose an iterative self-training method for hand segmentation in the new domain, which is guided by the model uncertainty estimated by a Bayesian CNN. We further use an adversarial component in our framework to utilize shared information about hand shapes to constrain the model adaptation process. Experiments on multiple egocentric datasets show that the proposed method significantly improves the generalization performance of hand segmentation.

A well trained and generalized deep neural network (DNN) should be robust to both seen and unseen classes. However, the performance of most of the existing supervised DNN algorithms degrade for classes which are unseen in the training set. To learn a discriminative classifier which yields good performance in Zero-Shot Learning (ZSL) settings, we propose to generate an Over-Complete Distribution (OCD) using Conditional Variational Autoencoder (CVAE) of both seen and unseen classes. In order to enforce the separability between classes and reduce the class scatter, we propose the use of Online Batch Triplet Loss (OBTL) and Center Loss (CL) on the generated OCD. The effectiveness of the framework is evaluated using both Zero-Shot Learning and Generalized Zero-Shot Learning protocols on three publicly available benchmark databases, SUN, CUB and AWA2. The results show that generating over-complete distributions and enforcing the classifier to learn a transform function from overlapping to non-overlapping distributions can improve the performance on both seen and unseen classes.

Image retrieval methods that employ hashing or vector quantization have achieved great success by taking advantage of deep learning. However, these approaches do not meet expectations unless expensive label information is sufficient. To resolve this issue, we propose the first quantization-based semi-supervised image retrieval scheme: Generalized Product Quantization (GPQ) network. We design a novel metric learning strategy that preserves semantic similarity between labeled data, and employ entropy regularization term to fully exploit inherent potentials of unlabeled data. Our solution increases the generalization capacity of the quantization network, which allows overcoming previous limitations in the retrieval community. Extensive experimental results demonstrate that GPQ yields state-of-the-art performance on large-scale real image benchmark datasets.

Deep neural networks have attained remarkable performance when applied to data that comes from the same distribution as that of the training set, but can significantly degrade otherwise. Therefore, detecting whether an example is out-of-distribution (OoD) is crucial to enable a system that can reject such samples or alert users. Recent works have made significant progress on OoD benchmarks consisting of small image datasets. However, many recent methods based on neural networks rely on training or tuning with both in-distribution and out-of-distribution data. The latter is generally hard to define a-priori, and its selection can easily bias the learning. We base our work on a popular method ODIN, proposing two strategies for freeing it from the needs of tuning with OoD data, while improving its OoD detection performance. We specifically propose to decompose confidence scoring as well as a modified input pre-processing method. We show that both of these significantly help in detection performance. Our further analysis on a larger scale image dataset shows that the two types of distribution shifts, specifically semantic shift and non-semantic shift, present a significant difference in the difficulty of the problem, providing an analysis of when ODIN-like strategies do or do not work.