CVPR 2020 Papers

We present an algorithm for extracting and vectorizing objects in images with polygons. Departing from a polygonal partition that oversegments an image into convex cells, the algorithm refines the geometry of the partition while labeling its cells by a semantic class. The result is a set of polygons, each capturing an object in the image. The quality of a configuration is measured by an energy that accounts for both the fidelity to input data and the complexity of the output polygons. To efficiently explore the configuration space, we perform splitting and merging operations in tandem on the cells of the polygonal partition. The exploration mechanism is controlled by a priority queue that sorts the operations most likely to decrease the energy. We show the potential of our algorithm on different types of scenes, from organic shapes to man-made objects through floor maps, and demonstrate its efficiency compared to existing vectorization methods.

We combine ideas from shock graph theory with more recent appearance-based methods for medial axis extraction from complex natural scenes, improving upon the present best unsupervised method, in terms of efficiency and performance. We make the following specific contributions: i) we extend the shock graph representation to the domain of real images, by generalizing the shock type definitions using local, appearance-based criteria; ii) we then use the rules of a Shock Grammar to guide our search for medial points, drastically reducing run time when compared to other methods, which exhaustively consider all points in the input image; iii) we remove the need for typical post-processing steps including thinning, non-maximum suppression, and grouping, by adhering to the Shock Grammar rules while deriving the medial axis solution; iv) finally, we raise some fundamental concerns with the evaluation scheme used in previous work and propose a more appropriate alternative for assessing the performance of medial axis extraction from scenes. Our experiments on the BMAX500 and SK-LARGE datasets demonstrate the effectiveness of our approach. We outperform the present state-of-the-art, excelling particularly in the high-precision regime, while running an order of magnitude faster and requiring no post-processing.

As a voxel-wise labeling task, semantic scene completion (SSC) tries to simultaneously infer the occupancy and semantic labels for a scene from a single depth and/or RGB image. The key challenge for SSC is how to effectively take advantage of the 3D context to model various objects or stuffs with severe variations in shapes, layouts, and visibility. To handle such variations, we propose a novel module called anisotropic convolution, which properties with flexibility and power impossible for the competing methods such as standard 3D convolution and some of its variations. In contrast to the standard 3D convolution that is limited to a fixed 3D receptive field, our module is capable of modeling the dimensional anisotropy voxel-wisely. The basic idea is to enable anisotropic 3D receptive field by decomposing a 3D convolution into three consecutive 1D convolutions, and the kernel size for each such 1D convolution is adaptively determined on the fly. By stacking multiple such anisotropic convolution modules, the voxel-wise modeling capability can be further enhanced while maintaining a controllable amount of model parameters. Extensive experiments on two SSC benchmarks, NYU-Depth-v2 and NYUCAD, show the superior performance of the proposed method.

Several studies show that animal needs are often expressed through their faces. Though remarkable progress has been made towards the automatic understanding of human faces, this has not been the case with animal faces. There exists significant room for algorithmic advances that could realize automatic systems for interpreting animal faces. Besides scientific value, resulting technology will foster better and cheaper animal care. We believe the underlying research progress is mainly obstructed by the lack of an adequately annotated dataset of animal faces, covering a wide spectrum of animal species. To this end, we introduce a large-scale, hierarchical annotated dataset of animal faces, featuring 22.4K faces from 350 diverse species and 21 animal orders across biological taxonomy. These faces are captured `in-the-wild' conditions and are consistently annotated with 9 landmarks on key facial features. The dataset is structured and scalable by design; its development underwent four systematic stages involving rigorous, overall effort of over 6K man-hours. We benchmark it for face alignment using the existing art under two new problem settings. Results showcase its challenging nature, unique attributes and present definite prospects for novel, adaptive, and generalized face-oriented CV algorithms. Further benchmarking the dataset across face detection and fine-grained recognition tasks demonstrates its multi-task applications and room for improvement. The dataset is available at: https://fdmaproject.wordpress.com/.

The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality.

Batch Normalization (BN) is extensively employed in various network architectures by performing standardization within mini-batches. A full understanding of the process has been a central target in the deep learning communities. Unlike existing works, which usually only analyze the standardization operation, this paper investigates the more general Batch Whitening (BW). Our work originates from the observation that while various whitening transformations equivalently improve the conditioning, they show significantly different behaviors in discriminative scenarios and training Generative Adversarial Networks (GANs). We attribute this phenomenon to the stochasticity that BW introduces. We quantitatively investigate the stochasticity of different whitening transformations and show that it correlates well with the optimization behaviors during training. We also investigate how stochasticity relates to the estimation of population statistics during inference. Based on our analysis, we provide a framework for designing and comparing BW algorithms in different scenarios. Our proposed BW algorithm improves the residual networks by a significant margin on ImageNet classification. Besides, we show that the stochasticity of BW can improve the GAN's performance with, however, the sacrifice of the training stability.

Batch Normalization (BN) techniques have been proposed to reduce the so-called Internal Covariate Shift (ICS) by attempting to keep the distributions of layer outputs unchanged. Experiments have shown their effectiveness on training deep neural networks. However, since only the first two moments are controlled in these BN techniques, it seems that a weak constraint is imposed on layer distributions and furthermore whether such constraint can reduce ICS is unknown. Thus this paper proposes a measure for ICS by using the Earth Mover (EM) distance and then derives the upper and lower bounds for the measure to provide a theoretical analysis of BN. The upper bound has shown that BN techniques can control ICS only for the outputs with low dimensions and small noise whereas their control is not effective in other cases. This paper also proves that such control is just a bounding of ICS rather than a reduction of ICS. Meanwhile, the analysis shows that the high-order moments and noise, which BN cannot control, have great impact on the lower bound. Based on such analysis, this paper furthermore proposes an algorithm that unitizes the outputs with an adjustable parameter to further bound ICS in order to cope with the problems of BN. The upper bound for the proposed unitization is noise-free and only dominated by the parameter. Thus, the parameter can be trained to tune the bound and further to control ICS. Besides, the unitization is embedded into the framework of BN to reduce the information loss. The experiments show that this proposed algorithm outperforms existing BN techniques on CIFAR-10, CIFAR-100 and ImageNet datasets.

Moving object segmentation in videos (MOS) is a highly demanding task for security-based applications like automated outdoor video surveillance. Most of the existing techniques proposed for MOS are highly depend on fine-tuning a model on the first frame(s) of test sequence or complicated training procedure, which leads to limited practical serviceability of the algorithm. In this paper, the inherent correlation learning-based edge extraction mechanism (EEM) and dense residual block (DRB) are proposed for the discriminative foreground representation. The multi-scale EEM module provides the efficient foreground edge related information (with the help of encoder) to the decoder through skip connection at subsequent scale. Further, the response of the optical flow encoder stream and the last EEM module are embedded in the bridge network. The bridge network comprises of multi-scale residual blocks with dense connections to learn the effective and efficient foreground relevant features. Finally, to generate accurate and consistent foreground object maps, a decoder block is proposed with skip connections from respective multi-scale EEM module feature maps and the subsequent down-sampled response of previous frame output. Specifically, the proposed network does not require any pre-trained models or fine-tuning of the parameters with the initial frame(s) of the test video. The performance of the proposed network is evaluated with different configurations like disjoint, cross-data, and global training-testing techniques. The ablation study is conducted to analyse each model of the proposed network. To demonstrate the effectiveness of the proposed framework, a comprehensive analysis on four benchmark video datasets is conducted. Experimental results show that the proposed approach outperforms the state-of-the-art methods for MOS.

Point clouds contain rich spatial information, which provides complementary cues for gesture recognition. In this paper, we formulate gesture recognition as an irregular sequence recognition problem and aim to capture long-term spatial correlations across point cloud sequences. A novel and effective PointLSTM is proposed to propagate information from past to future while preserving the spatial structure. The proposed PointLSTM combines state information from neighboring points in the past with current features to update the current states by a weight-shared LSTM layer. This method can be integrated into many other sequence learning approaches. In the task of gesture recognition, the proposed PointLSTM achieves state-of-the-art results on two challenging datasets (NVGesture and SHREC'17) and outperforms previous skeleton-based methods. To show its advantages in generalization, we evaluate our method on MSR Action3D dataset, and it produces competitive results with previous skeleton-based methods.

Automatically finding suspicious regions in a potentially forged image by splicing, inpainting or copy-move remains a widely open problem. Blind detection neural networks trained on benchmark data are flourishing. Yet, these methods do not provide an explanation of their detections. The more traditional methods try to provide such evidence by pointing out local inconsistencies in the image noise, JPEG compression, chromatic aberration, or in the mosaic. In this paper we develop a blind method that can train directly on unlabelled and potentially forged images to point out local mosaic inconsistencies. To this aim we designed a CNN structure inspired from demosaicing algorithms and directed at classifying image blocks by their position in the image modulo (2 x 2). Creating a diversified benchmark database using varied demosaicing methods, we explore the efficiency of the method and its ability to adapt quickly to any new data.

In order to alleviate the notorious mode collapse phenomenon in generative adversarial networks (GANs), we propose a novel training method of GANs in which certain fake samples are considered as real ones during the training process. This strategy can reduce the gradient value that generator receives in the region where gradient exploding happens. We show the process of an unbalanced generation and a vicious circle issue resulted from gradient exploding in practical training, which explains the instability of GANs. We also theoretically prove that gradient exploding can be alleviated by penalizing the difference between discriminator outputs and fake-as-real consideration for very close real and fake samples. Accordingly, Fake-As-Real GAN (FARGAN) is proposed with a more stable training process and a more faithful generated distribution. Experiments on different datasets verify our theoretical analysis.

Many methods have set state-of-the-art performance on restoring images degraded by bad weather such as rain, haze, fog, and snow, however they are designed specifically to handle one type of degradation. In this paper, we propose a method that can handle multiple bad weather degradations: rain, fog, snow and adherent raindrops using a single network. To achieve this, we first design a generator with multiple task-specific encoders, each of which is associated with a particular bad weather degradation type. We utilize a neural architecture search to optimally process the image features extracted from all encoders. Subsequently, to convert degraded image features to clean background features, we introduce a series of tensor-based operations encapsulating the underlying physics principles behind the formation of rain, fog, snow and adherent raindrops. These operations serve as the basic building blocks for our architectural search. Finally, our discriminator simultaneously assesses the correctness and classifies the degradation type of the restored image. We design a novel adversarial learning scheme that only backpropagates the loss of a degradation type to the respective task-specific encoder. Despite being designed to handle different types of bad weather, extensive experiments demonstrate that our method performs competitively to the individual and dedicated state-of-the-art image restoration methods.

The success of deep learning in visual recognition tasks has driven advancements in multiple fields of research. Particularly, increasing attention has been drawn towards its application in agriculture. Nevertheless, while visual pattern recognition on farmlands carries enormous economic values, little progress has been made to merge computer vision and crop sciences due to the lack of suitable agricultural image datasets. Meanwhile, problems in agriculture also pose new challenges in computer vision. For example, semantic segmentation of aerial farmland images requires inference over extremely large-size images with extreme annotation sparsity. These challenges are not present in most of the common object datasets, and we show that they are more challenging than many other aerial image datasets. To encourage research in computer vision for agriculture, we present Agriculture-Vision: a large-scale aerial farmland image dataset for semantic segmentation of agricultural patterns. We collected 94,986 high-quality aerial images from 3,432 farmlands across the US, where each image consists of RGB and Near-infrared (NIR) channels with resolution as high as 10 cm per pixel. We annotate nine types of field anomaly patterns that are most important to farmers. As a pilot study of aerial agricultural semantic segmentation, we perform comprehensive experiments using popular semantic segmentation models; we also propose an effective model designed for aerial agricultural pattern recognition. Our experiments demonstrate several challenges Agriculture-Vision poses to both the computer vision and agriculture communities. Future versions of this dataset will include even more aerial images, anomaly patterns and image channels.

Affinity graphs are widely used in deep architectures, including graph convolutional neural networks and attention networks. Thus far, the literature has focused on abstracting features from such graphs, while the learning of the affinities themselves has been overlooked. Here we propose a principled method to directly supervise the learning of weights in affinity graphs, to exploit meaningful connections between entities in the data source. Applied to a visual attention network, our affinity supervision improves relationship recovery between objects, even without the use of manually annotated relationship labels. We further show that affinity learning between objects boosts scene categorization performance and that the supervision of affinity can also be applied to graphs built from mini-batches, for neural network training. In an image classification task we demonstrate consistent improvement over the baseline, with diverse network architectures and datasets.

Humans learn to drive through both practice and theory, e.g. by studying the rules, while most self-driving systems are limited to the former. Being able to incorporate human knowledge of typical causal driving behaviour should benefit autonomous systems. We propose a new approach that learns vehicle control with the help of human advice. Specifically, our system learns to summarize its visual observations in natural language, predict an appropriate action response (e.g. "I see a pedestrian crossing, so I stop"), and predict the controls, accordingly. Moreover, to enhance interpretability of our system, we introduce a fine-grained attention mechanism which relies on semantic segmentation and object-centric RoI pooling. We show that our approach of training the autonomous system with human advice, grounded in a rich semantic representation, matches or outperforms prior work in terms of control prediction and explanation generation. Our approach also results in more interpretable visual explanations by visualizing object-centric attention maps. Code is available at https://github.com/JinkyuKimUCB/advisable-driving.

Neural Architecture Search (NAS) that aims to automate the procedure of architecture design has achieved promising results in many computer vision fields. In this paper, we propose an AdversarialNAS method specially tailored for Generative Adversarial Networks (GANs) to search for a superior generative model on the task of unconditional image generation. The AdversarialNAS is the first method that can search the architectures of generator and discriminator simultaneously in a differentiable manner. During searching, the designed adversarial search algorithm does not need to comput any extra metric to evaluate the performance of the searched architecture, and the search paradigm considers the relevance between the two network architectures and improves their mutual balance. Therefore, AdversarialNAS is very efficient and only takes 1 GPU day to search for a superior generative model in the proposed large search space. Experiments demonstrate the effectiveness and superiority of our method. The discovered generative model sets a new state-of-the-art FID score of 10.87 and highly competitive Inception Score of 8.74 on CIFAR-10. Its transferability is also proven by setting new state-of-the-art FID score of 26.98 and Inception score of 9.63 on STL-10. Code is at: https://github.com/chengaopro/AdversarialNAS.

Adversarial examples cause neural networks to produce incorrect outputs with high confidence. Although adversarial training is one of the most effective forms of defense against adversarial examples, unfortunately, a large gap exists between test accuracy and training accuracy in adversarial training. In this paper, we identify Adversarial Feature Overfitting (AFO), which may cause poor adversarially robust generalization, and we show that adversarial training can overshoot the optimal point in terms of robust generalization, leading to AFO in our simple Gaussian model. Considering these theoretical results, we present soft labeling as a solution to the AFO problem. Furthermore, we propose Adversarial Vertex mixup (AVmixup), a soft-labeled data augmentation approach for improving adversarially robust generalization. We complement our theoretical analysis with experiments on CIFAR10, CIFAR100, SVHN, and Tiny ImageNet, and show that AVmixup significantly improves the robust generalization performance and that it reduces the trade-off between standard accuracy and adversarial robustness.

Realistic color texture generation is an important step in RGB-D surface reconstruction, but remains challenging in practice due to inaccuracies in reconstructed geometry, misaligned camera poses, and view-dependent imaging artifacts. In this work, we present a novel approach for color texture generation using a conditional adversarial loss obtained from weakly-supervised views. Specifically, we propose an approach to produce photorealistic textures for approximate surfaces, even from misaligned images, by learning an objective function that is robust to these errors. The key idea of our approach is to learn a patch-based conditional discriminator which guides the texture optimization to be tolerant to misalignments. Our discriminator takes a synthesized view and a real image, and evaluates whether the synthesized one is realistic, under a broadened definition of realism. We train the discriminator by providing as `real' examples pairs of input views and their misaligned versions -- so that the learned adversarial loss will tolerate errors from the scans. Experiments on synthetic and real data under quantitative or qualitative evaluation demonstrate the advantage of our approach in comparison to state of the art.

Pretrained models from self-supervision are prevalently used in fine-tuning downstream tasks faster or for better accuracy. However, gaining robustness from pretraining is left unexplored. We introduce adversarial training into self-supervision, to provide general-purpose robust pretrained models for the first time. We find these robust pretrained models can benefit the subsequent fine-tuning in two ways: i) boosting final model robustness; ii) saving the computation cost, if proceeding towards adversarial fine-tuning. We conduct extensive experiments to demonstrate that the proposed framework achieves large performance margins (eg, 3.83% on robust accuracy and 1.3% on standard accuracy, on the CIFAR-10 dataset), compared with the conventional end-to-end adversarial training baseline. Moreover, we find that different self-supervised pretrained models have diverse adversarial vulnerability. It inspires us to ensemble several pretraining tasks, which boosts robustness more. Our ensemble strategy contributes to a further improvement of 3.59% on robust accuracy, while maintaining a slightly higher standard accuracy on CIFAR-10. Our codes are available at https://github.com/TAMU-VITA/Adv-SS-Pretraining.