ICCV 2019 Papers

Image extension models have broad applications in image editing, computational photography and computer graphics. While image inpainting has been extensively studied in the literature, it is challenging to directly apply the state-of-the-art inpainting methods to image extension as they tend to generate blurry or repetitive pixels with inconsistent semantics. We introduce semantic conditioning to the discriminator of a generative adversarial network (GAN), and achieve strong results on image extension with coherent semantics and visually pleasing colors and textures. We also show promising results in extreme extensions, such as panorama generation.

In this work, we address the challenging issue of scene segmentation. To increase the feature similarity of the same object while keeping the feature discrimination of different objects, we explore to propagate information throughout the image under the control of objects' boundaries. To this end, we first propose to learn the boundary as an additional semantic class to enable the network to be aware of the boundary layout. Then, we propose unidirectional acyclic graphs (UAGs) to model the function of undirected cyclic graphs (UCGs), which structurize the image via building graphic pixel-by-pixel connections, in an efficient and effective way. Furthermore, we propose a boundary-aware feature propagation (BFP) module to harvest and propagate the local features within their regions isolated by the learned boundaries in the UAG-structured image. The proposed BFP is capable of splitting the feature propagation into a set of semantic groups via building strong connections among the same segment region but weak connections between different segment regions. Without bells and whistles, our approach achieves new state-of-the-art segmentation performance on three challenging semantic segmentation datasets, i.e., PASCAL-Context, CamVid, and Cityscapes.

We consider general discrete Markov Random Fields(MRFs) with additional bottleneck potentials which penalize the maximum (instead of the sum) over local potential value taken by the MRF-assignment. Bottleneck potentials or analogous constructions have been considered in (i) combinatorial optimization (e.g. bottleneck shortest path problem, the minimum bottleneck spanning tree problem, bottleneck function minimization in greedoids), (ii) inverse problems with L_ infinity -norm regularization and (iii) valued constraint satisfaction on the (min,max)-pre-semirings. Bottleneck potentials for general discrete MRFs are a natural generalization of the above direction of modeling work to Maximum-A-Posteriori (MAP) inference in MRFs. To this end we propose MRFs whose objective consists of two parts: terms that factorize according to (i) (min,+), i.e. potentials as in plain MRFs, and (ii) (min,max), i.e. bottleneck potentials. To solve the ensuing inference problem, we propose high-quality relaxations and efficient algorithms for solving them. We empirically show efficacy of our approach on large scale seismic horizon tracking problems.

Few-shot learning and self-supervised learning address different facets of the same problem: how to train a model with little or no labeled data. Few-shot learning aims for optimization methods and models that can learn efficiently to recognize patterns in the low data regime. Self-supervised learning focuses instead on unlabeled data and looks into it for the supervisory signal to feed high capacity deep neural networks. In this work we exploit the complementarity of these two domains and propose an approach for improving few-shot learning through self-supervision. We use self-supervision as an auxiliary task in a few-shot learning pipeline, enabling feature extractors to learn richer and more transferable visual representations while still using few annotated samples. Through self-supervision, our approach can be naturally extended towards using diverse unlabeled data from other datasets in the few-shot setting. We report consistent improvements across an array of architectures, datasets and self-supervision techniques. We provide the implementation code at: https://github.com/valeoai/BF3S

Image datasets with high-quality pixel-level annotations are valuable for semantic segmentation: labelling every pixel in an image ensures that rare classes and small objects are annotated. However, full-image annotations are expensive, with experts spending up to 90 minutes per image. We propose block sub-image annotation as a replacement for full-image annotation. Despite the attention cost of frequent task switching, we find that block annotations can be crowdsourced at higher quality compared to full-image annotation with equal monetary cost using existing annotation tools developed for full-image annotation. Surprisingly, we find that 50% pixels annotated with blocks allows semantic segmentation to achieve equivalent performance to 100% pixels annotated. Furthermore, as little as 12% of pixels annotated allows performance as high as 98% of the performance with dense annotation. In weakly-supervised settings, block annotation outperforms existing methods by 3-4% (absolute) given equivalent annotation time. To recover the necessary global structure for applications such as characterizing spatial context and affordance relationships, we propose an effective method to inpaint block-annotated images with high-quality labels without additional human effort. As such, fewer annotations can also be used for these applications compared to full-image annotation.

Several important security issues of Deep Neural Network (DNN) have been raised recently associated with different applications and components. The most widely investigated security concern of DNN is from its malicious input, a.k.a adversarial example. Nevertheless, the security challenge of DNN's parameters is not well explored yet. In this work, we are the first to propose a novel DNN weight attack methodology called Bit-Flip Attack (BFA) which can crush a neural network through maliciously flipping extremely small amount of bits within its weight storage memory system (i.e., DRAM). The bit-flip operations could be conducted through well-known Row-Hammer attack, while our main contribution is to develop an algorithm to identify the most vulnerable bits of DNN weight parameters (stored in memory as binary bits), that could maximize the accuracy degradation with a minimum number of bit-flips. Our proposed BFA utilizes a Progressive Bit Search (PBS) method which combines gradient ranking and progressive search to identify the most vulnerable bit to be flipped. With the aid of PBS, we can successfully attack a ResNet-18 fully malfunction (i.e., top-1 accuracy degrade from 69.8% to 0.1%) only through 13 bit-flips out of 93 million bits, while randomly flipping 100 bits merely degrades the accuracy by less than 1%. Code is released at: https://github.com/elliothe/Neural_Network_Weight_Attack

This paper investigates a novel Bilinear attention (Bi-attention) block, which discovers and uses second order statistical information in an input feature map, for the purpose of person retrieval. The Bi-attention block uses bilinear pooling to model the local pairwise feature interactions along each channel, while preserving the spatial structural information. We propose an Attention in Attention (AiA) mechanism to build inter-dependency among the second order local and global features with the intent to make better use of, or pay more attention to, such higher order statistical relationships. The proposed network, equipped with the proposed Bi-attention is referred to as Bilinear ATtention network (BAT-net). Our approach outperforms current state-of-the-art by a considerable margin across the standard benchmark datasets (e.g., CUHK03, Market-1501, DukeMTMC-reID and MSMT17).

In this paper, we study fast training of adversarially robust models. From the analyses of the state-of-the-art defense method, i.e., the multi-step adversarial training [??], we hypothesize that the gradient magnitude links to the model robustness. Motivated by this, we propose to perturb both the image and the label during training, which we call Bilateral Adversarial Training (BAT). To generate the adversarial label, we derive an closed-form heuristic solution. To generate the adversarial image, we use one-step targeted attack with the target label being the most confusing class. In the experiment, we first show that random start and the most confusing target attack effectively prevent the label leaking and gradient masking problem. Then coupled with the adversarial label part, our model significantly improves the state-of-the-art results. For example, against PGD100 white-box attack with cross-entropy loss, on CIFAR10, we achieve 63.7% versus 47.2%; on SVHN, we achieve 59.1% versus 42.1%. At last, the experiment on the very (computationally) challenging ImageNet dataset further demonstrates the effectiveness of our fast method.

We study the problem of mapping between a domain A, in which there is a single training sample and a domain B, for which we have a richer training set. The method we present is able to perform this mapping in both directions. For example, we can transfer all MNIST images to the visual domain captured by a single SVHN image and transform the SVHN image to the domain of the MNIST images. Our method is based on employing one encoder and one decoder for each domain, without utilizing weight sharing. The autoencoder of the single sample domain is trained to match both this sample and the latent space of domain B. Our results demonstrate convincing mapping between domains, where either the source or the target domain are defined by a single sample, far surpassing existing solutions. Our code is made publicly available at https://github.com/tomercohen11/BiOST.

Person re-identification is a challenging task due to various complex factors. Recent studies have attempted to integrate human parsing results or externally defined attributes to help capture human parts or important object regions. On the other hand, there still exist many useful contextual cues that do not fall into the scope of predefined human parts or attributes. In this paper, we address the missed contextual cues by exploiting both the accurate human parts and the coarse non-human parts. In our implementation, we apply a human parsing model to extract the binary human part masks and a self-attention mechanism to capture the soft latent (non-human) part masks. We verify the effectiveness of our approach with new state-of-the-art performance on three challenging benchmarks: Market-1501, DukeMTMC-reID and CUHK03. Our implementation is available at https://github.com/ggjy/P2Net.pytorch.

The dominant approach for learning local patch descriptors relies on small image regions whose scale must be properly estimated a priori by a keypoint detector. In other words, if two patches are not in correspondence, their descriptors will not match. A strategy often used to alleviate this problem is to "pool" the pixel-wise features over log-polar regions, rather than regularly spaced ones. By contrast, we propose to extract the "support region" directly with a log-polar sampling scheme. We show that this provides us with a better representation by simultaneously oversampling the immediate neighbourhood of the point and undersampling regions far away from it. We demonstrate that this representation is particularly amenable to learning descriptors with deep networks. Our models can match descriptors across a much wider range of scales than was possible before, and also leverage much larger support regions without suffering from occlusions. We report state-of-the-art results on three different datasets

In spite of recent success of proposal-based CNN models for object detection, it is still difficult to detect small objects due to the limited and distorted information that small region of interests (RoI) contain. One way to alleviate this issue is to enhance the features of small RoIs using a super-resolution (SR) technique. We investigate how to improve feature-level super-resolution especially for small object detection, and discover its performance can be significantly improved by (i) utilizing proper high-resolution target features as supervision signals for training of a SR model and (ii) matching the relative receptive fields of training pairs of input low-resolution features and target high-resolution features. We propose a novel feature-level super-resolution approach that not only correctly addresses these two desiderata but also is integrable with any proposal-based detectors with feature pooling. In our experiments, our approach significantly improves the performance of Faster R-CNN on three benchmarks of Tsinghua-Tencent 100K, PASCAL VOC and MS COCO. The improvement for small objects is remarkably large, and encouragingly, those for medium and large objects are nontrivial too. As a result, we achieve new state-of-the-art performance on Tsinghua-Tencent 100K and highly competitive results on both PASCAL VOC and MS COCO.

Recent studies on image patch matching are paying more attention on hard sample learning, because easy samples do not contribute much to the network optimization. They have proposed various hard negative sample mining strategies, but very few addressed this problem from the perspective of loss functions. Our research shows that the conventional Siamese and triplet losses treat all samples linearly, thus make the training time consuming. Instead, we propose the exponential Siamese and triplet losses, which can naturally focus more on hard samples and put less emphasis on easy ones, meanwhile, speed up the optimization. To assist the exponential losses, we introduce the hard positive sample mining to further enhance the effectiveness. The extensive experiments demonstrate our proposal improves both metric and descriptor learning on several well accepted benchmarks, and outperforms the state-of-the-arts on the UBC dataset. Moreover, it also shows a better generalizability on cross-spectral image matching and image retrieval tasks.

Convolutional neural networks have been widely deployed in various application scenarios. In order to extend the applications' boundaries to some accuracy-crucial domains, researchers have been investigating approaches to boost accuracy through either deeper or wider network structures, which brings with them the exponential increment of the computational and storage cost, delaying the responding time. In this paper, we propose a general training framework named self distillation, which notably enhances the performance (accuracy) of convolutional neural networks through shrinking the size of the network rather than aggrandizing it. Different from traditional knowledge distillation - a knowledge transformation methodology among networks, which forces student neural networks to approximate the softmax layer outputs of pre-trained teacher neural networks, the proposed self distillation framework distills knowledge within network itself. The networks are firstly divided into several sections. Then the knowledge in the deeper portion of the networks is squeezed into the shallow ones. Experiments further prove the generalization of the proposed self distillation framework: enhancement of accuracy at average level is 2.65%, varying from 0.61% in ResNeXt as minimum to 4.07% in VGG19 as maximum. In addition, it can also provide flexibility of depth-wise scalable inference on resource-limited edge devices. Our codes have been released on github.

We introduce a new memory architecture, Bayesian Relational Memory (BRM), to improve the generalization ability for semantic visual navigation agents in unseen environments, where an agent is given a semantic target to navigate towards. BRM takes the form of a probabilistic relation graph over semantic entities (e.g., room types), which allows (1) capturing the layout prior from training environments, i.e., prior knowledge, (2) estimating posterior layout at test time, i.e., memory update, and (3) efficient planning for navigation, altogether. We develop a BRM agent consisting of a BRM module for producing sub-goals and a goal-conditioned locomotion module for control. When testing in unseen environments, the BRM agent outperforms baselines that do not explicitly utilize the probabilistic relational memory structure.

Deep convolutional neural networks (DCNNs) have dominated the recent developments in computer vision through making various record-breaking models. However, it is still a great challenge to achieve powerful DCNNs in resource-limited environments, such as on embedded devices and smart phones. Researchers have realized that 1-bit CNNs can be one feasible solution to resolve the issue; however, they are baffled by the inferior performance compared to the full-precision DCNNs. In this paper, we propose a novel approach, called Bayesian optimized 1-bit CNNs (denoted as BONNs), taking the advantage of Bayesian learning, a well-established strategy for hard problems, to significantly improve the performance of extreme 1-bit CNNs. We incorporate the prior distributions of full-precision kernels and features into the Bayesian framework to construct 1-bit CNNs in an end-to-end manner, which have not been considered in any previous related methods. The Bayesian losses are achieved with a theoretical support to optimize the network simultaneously in both continuous and discrete spaces, aggregating different losses jointly to improve the model capacity. Extensive experiments on the ImageNet and CIFAR datasets show that BONNs achieve the best classification performance compared to state-of-the-art 1-bit CNNs.

In crowd counting datasets, each person is annotated by a point, which is usually the center of the head. And the task is to estimate the total count in a crowd scene. Most of the state-of-the-art methods are based on density map estimation, which convert the sparse point annotations into a "ground truth" density map through a Gaussian kernel, and then use it as the learning target to train a density map estimator. However, such a "ground-truth" density map is imperfect due to occlusions, perspective effects, variations in object shapes, etc. On the contrary, we propose Bayesian loss, a novel loss function which constructs a density contribution probability model from the point annotations. Instead of constraining the value at every pixel in the density map, the proposed training loss adopts a more reliable supervision on the count expectation at each annotated point. Without bells and whistles, the loss function makes substantial improvements over the baseline loss on all tested datasets. Moreover, our proposed loss function equipped with a standard backbone network, without using any external detectors or multi-scale architectures, plays favourably against the state of the arts. Our method outperforms previous best approaches by a large margin on the latest and largest UCF-QNRF dataset.

We propose a framework for recognizing human actions from skeleton data by modeling the underlying dynamic process that generates the motion pattern. We capture three major factors that contribute to the complexity of the motion pattern including spatial dependencies among body joints, temporal dependencies of body poses, and variation among subjects in action execution. We utilize graph convolution to extract structure-aware feature representation from pose data by exploiting the skeleton anatomy. Long short-term memory (LSTM) network is then used to capture the temporal dynamics of the data. Finally, the whole model is extended under the Bayesian framework to a probabilistic model in order to better capture the stochasticity and variation in the data. An adversarial prior is developed to regularize the model parameters to improve the generalization of the model. A Bayesian inference problem is formulated to solve the classification task. We demonstrate the benefit of this framework in several benchmark datasets with recognition under various generalization conditions.

Superpixels provide a useful intermediate image representation. Existing superpixel methods, however, suffer from at least some of the following drawbacks: 1) topology is handled heuristically; 2) the number of superpixels is either predefined or estimated at a prohibitive cost; 3) lack of adaptiveness. As a remedy, we propose a novel probabilistic model, self-coined Bayesian Adaptive Superpixel Segmentation (BASS), together with an efficient inference. BASS is a Bayesian nonparametric mixture model that also respects topology and favors spatial coherence. The optimizationbased and topology-aware inference is parallelizable and implemented in GPU. Quantitatively, BASS achieves results that are either better than the state-of-the-art or close to it, depending on the performance index and/or dataset. Qualitatively, we argue it achieves the best results; we demonstrate this by not only subjective visual inspection but also objective quantitative performance evaluation of the downstream application of face detection. Our code is available at https://github.com/uzielroy/BASS.