2018-11-20
这篇文章介绍46篇论文，涉及CNN、Face、图像分类、目标检测、图像分割、GAN、Re-ID、SLAM和迁移学习等方向。
《Deeper Interpretability of Deep Networks》
arXiv：https://arxiv.org/abs/1811.07807
Deep Convolutional Neural Networks (CNNs) have been one of the most influential recent developments in computer vision, particularly for categorization. There is an increasing demand for explainable AI as these systems are deployed in the real world. However, understanding the information represented and processed in CNNs remains in most cases challenging. Within this paper, we explore the use of new information theoretic techniques developed in the field of neuroscience to enable novel understanding of how a CNN represents information. We trained a 10-layer ResNet architecture to identify 2,000 face identities from 26M images generated using a rigorously controlled 3D face rendering model that produced variations of intrinsic (i.e. face morphology, gender, age, expression and ethnicity) and extrinsic factors (i.e. 3D pose, illumination, scale and 2D translation). With our methodology, we demonstrate that unlike human's network overgeneralizes face identities even with extreme changes of face shape, but it is more sensitive to changes of texture. To understand the processing of information underlying these counterintuitive properties, we visualize the features of shape and texture that the network processes to identify faces. Then, we shed a light into the inner workings of the black box and reveal how hidden layers represent these features and whether the representations are invariant to pose. We hope that our methodology will provide an additional valuable tool for interpretability of CNNs.
《Deep Shape-from-Template: Wide-Baseline, Dense and Fast Registration and Deformable Reconstruction from a Single Image》
arXiv：https://arxiv.org/abs/1811.07791
We present Deep Shape-from-Template (DeepSfT), a novel Deep Neural Network (DNN) method for solving real-time automatic registration and 3D reconstruction of a deformable object viewed in a single monocular image.DeepSfT advances the state-of-the-art in various aspects. Compared to existing DNN SfT methods, it is the first fully convolutional real-time approach that handles an arbitrary object geometry, topology and surface representation. It also does not require ground truth registration with real data and scales well to very complex object models with large numbers of elements. Compared to previous non-DNN SfT methods, it does not involve numerical optimization at run-time, and is a dense, wide-baseline solution that does not demand, and does not suffer from, feature-based matching. It is able to process a single image with significant deformation and viewpoint changes, and handles well the core challenges of occlusions, weak texture and blur. DeepSfT is based on residual encoder-decoder structures and refining blocks. It is trained end-to-end with a novel combination of supervised learning from simulated renderings of the object model and semi-supervised automatic fine-tuning using real data captured with a standard RGB-D camera. The cameras used for fine-tuning and run-time can be different, making DeepSfT practical for real-world use. We show that DeepSfT significantly outperforms state-of-the-art wide-baseline approaches for non-trivial templates, with quantitative and qualitative evaluation.
《Do Normalization Layers in a Deep ConvNet Really Need to Be Distinct?》
arXiv：https://arxiv.org/abs/1811.07727
Yes, they do. This work investigates a perspective for deep learning: whether different normalization layers in a ConvNet require different normalizers. This is the first step towards understanding this phenomenon. We allow each convolutional layer to be stacked before a switchable normalization (SN) that learns to choose a normalizer from a pool of normalization methods. Through systematic experiments in ImageNet, COCO, Cityscapes, and ADE20K, we answer three questions: (a) Is it useful to allow each normalization layer to select its own normalizer? (b) What impacts the choices of normalizers? (c) Do different tasks and datasets prefer different normalizers? Our results suggest that (1) using distinct normalizers improves both learning and generalization of a ConvNet; (2) the choices of normalizers are more related to depth and batch size, but less relevant to parameter initialization, learning rate decay, and solver; (3) different tasks and datasets have different behaviors when learning to select normalizers.
《Self-Referenced Deep Learning》
arXiv：https://arxiv.org/abs/1811.07598
Knowledge distillation is an effective approach to transferring knowledge from a teacher neural network to a student target network for satisfying the low-memory and fast running requirements in practice use. Whilst being able to create stronger target networks compared to the vanilla non-teacher based learning strategy, this scheme needs to train additionally a large teacher model with expensive computational cost. In this work, we present a Self-Referenced Deep Learning (SRDL) strategy. Unlike both vanilla optimisation and existing knowledge distillation, SRDL distils the knowledge discovered by the in-training target model back to itself to regularise the subsequent learning procedure therefore eliminating the need for training a large teacher model. SRDL improves the model generalisation performance compared to vanilla learning and conventional knowledge distillation approaches with negligible extra computational cost. Extensive evaluations show that a variety of deep networks benefit from SRDL resulting in enhanced deployment performance on both coarse-grained object categorisation tasks (CIFAR10, CIFAR100, Tiny ImageNet, and ImageNet) and fine-grained person instance identification tasks (Market-1501).
《Multimodal Densenet》
arXiv：https://arxiv.org/abs/1811.07407
Humans make accurate decisions by interpreting complex data from multiple sources. Medical diagnostics, in particular, often hinge on human interpretation of multi-modal information. In order for artificial intelligence to make progress in automated, objective, and accurate diagnosis and prognosis, methods to fuse information from multiple medical imaging modalities are required. However, combining information from multiple data sources has several challenges, as current deep learning architectures lack the ability to extract useful representations from multimodal information, and often simple concatenation is used to fuse such information. In this work, we propose Multimodal DenseNet, a novel architecture for fusing multimodal data. Instead of focusing on concatenation or early and late fusion, our proposed architectures fuses information over several layers and gives the model flexibility in how it combines information from multiple sources. We apply this architecture to the challenge of polyp characterization and landmark identification in endoscopy. Features from white light images are fused with features from narrow band imaging or depth maps. This study demonstrates that Multimodal DenseNet outperforms monomodal classification as well as other multimodal fusion techniques by a significant margin on two different datasets.
《RePr: Improved Training of Convolutional Filters》
arXiv：https://arxiv.org/abs/1811.07275
A well-trained Convolutional Neural Network can easily be pruned without significant loss of performance. This is because of unnecessary overlap in the features captured by the network's filters. Innovations in network architecture such as skip/dense connections and Inception units have mitigated this problem to some extent, but these improvements come with increased computation and memory requirements at run-time. We attempt to address this problem from another angle - not by changing the network structure but by altering the training method. We show that by temporarily pruning and then restoring a subset of the model's filters, and repeating this process cyclically, overlap in the learned features is reduced, producing improved generalization. We show that the existing model-pruning criteria are not optimal for selecting filters to prune in this context and introduce inter-filter orthogonality as the ranking criteria to determine under-expressive filters. Our method is applicable both to vanilla convolutional networks and more complex modern architectures, and improves the performance across a variety of tasks, especially when applied to smaller networks.
《PydMobileNet: Improved Version of MobileNets with Pyramid Depthwise Separable Convolution》
arXiv：https://arxiv.org/abs/1811.07083
Convolutional neural networks (CNNs) have shown remarkable performance in various computer vision tasks in recent years. However, the increasing model size has raised challenges in adopting them in real-time applications as well as mobile and embedded vision applications. Many works try to build networks as small as possible while still have acceptable performance. The state-of-the-art architecture is MobileNets. They use Depthwise Separable Convolution (DWConvolution) in place of standard Convolution to reduce the size of networks. This paper describes an improved version of MobileNet, called Pyramid Mobile Network. Instead of using just a 3×3 kernel size for DWConvolution like in MobileNet, the proposed network uses a pyramid kernel size to capture more spatial information. The proposed architecture is evaluated on two highly competitive object recognition benchmark datasets (CIFAR-10, CIFAR-100). The experiments demonstrate that the proposed network achieves better performance compared with MobileNet as well as other state-of-the-art networks. Additionally, it is more flexible in fine-tuning the trade-off between accuracy, latency and model size than MobileNets.
《Aff-Wild2: Extending the Aff-Wild Database for Affect Recognition》
arXiv：https://arxiv.org/abs/1811.07770
Automatic understanding of human affect using visual signals is a problem that has attracted significant interest over the past 20 years. However, human emotional states are quite complex. To appraise such states displayed in real-world settings, we need expressive emotional descriptors that are capable of capturing and describing this complexity. The circumplex model of affect, which is described in terms of valence (i.e., how positive or negative is an emotion) and arousal (i.e., power of the activation of the emotion), can be used for this purpose. Recent progress in the emotion recognition domain has been achieved through the development of deep neural architectures and the availability of very large training databases. To this end, Aff-Wild has been the first large-scale "in-the-wild" database, containing around 1,200,000 frames. In this paper, we build upon this database, extending it with 260 more subjects and 1,413,000 new video frames. We call the union of Aff-Wild with the additional data, Aff-Wild2. The videos are downloaded from Youtube and have large variations in pose, age, illumination conditions, ethnicity and profession. Both database-specific as well as cross-database experiments are performed in this paper, by utilizing the Aff-Wild2, along with the RECOLA database. The developed deep neural architectures are based on the joint training of state-of-the-art convolutional and recurrent neural networks with attention mechanism; thus exploiting both the invariant properties of convolutional features, while modeling temporal dynamics that arise in human behaviour via the recurrent layers. The obtained results show premise for utilization of the extended Aff-Wild, as well as of the developed deep neural architectures for visual analysis of human behaviour in terms of continuous emotion dimensions.
《High Order Neural Networks for Video Classification》
arXiv：https://arxiv.org/abs/1811.07519
Capturing spatiotemporal correlations is an essential topic in video classification. In this paper, we present high order operations as a generic family of building blocks for capturing high order correlations from high dimensional input video space. We prove that several successful architectures for visual classification tasks are in the family of high order neural networks, theoretical and experimental analysis demonstrates their underlying mechanism is high order. We also proposal a new LEarnable hiGh Order (LEGO) block, whose goal is to capture spatiotemporal correlation in a feedforward manner. Specifically, LEGO blocks implicitly learn the relation expressions for spatiotemporal features and use the learned relations to weight input features. This building block can be plugged into many neural network architectures, achieving evident improvement without introducing much overhead. On the task of video classification, even using RGB only without fine-tuning with other video datasets, our high order models can achieve results on par with or better than the existing state-of-the-art methods on both Something-Something (V1 and V2) and Charades datasets.
《DeepConsensus: using the consensus of features from multiple layers to attain robust image classification》
arXiv：https://arxiv.org/abs/1811.07266
We consider a classifier whose test set is exposed to various perturbations that are not present in the training set. These test samples still contain enough features to map them to the same class as their unperturbed counterpart. Current architectures exhibit rapid degradation of accuracy when trained on standard datasets but then used to classify perturbed samples of that data. To address this, we present a novel architecture named DeepConsensus that significantly improves generalization to these test-time perturbations. Our key insight is that deep neural networks should directly consider summaries of low and high level features when making classifications. Existing convolutional neural networks can be augmented with DeepConsensus, leading to improved resistance against large and small perturbations on MNIST, EMNIST, FashionMNIST, CIFAR10 and SVHN datasets.
《Weakly Supervised Soft-detection-based Aggregation Method for Image Retrieval》
arXiv：https://arxiv.org/abs/1811.07619
In recent year, the compact representations based on activations of Convolutional Neural Network (CNN) achieve remarkable performance in image retrieval. Some interested object only takes up a small part of the whole image. Therefore, it is significant to extract the discriminative representations that contain regional information of pivotal small object. In this paper, we propose a novel weakly supervised soft-detection-based aggregation (SDA) method free from bounding box annotations for image retrieval. In order to highlight the certain discriminative pattern of objects and suppress the noise of background, we employ trainable soft region proposals that indicate the probability of interested object and reflect the significance of candidate regions. We conduct comprehensive experiments on standard image retrieval datasets. Our weakly supervised SDA method achieves state-of-the-art performance on most benchmarks. The results demonstrate that the proposed SDA method is effective for image retrieval.
《Fast Efficient Object Detection Using Selective Attention》
arXiv：https://arxiv.org/abs/1811.07502
Deep learning object detectors achieve state-of-the-art accuracy at the expense of high computational overheads, impeding their utilization on embedded systems such as drones. A primary source of these overheads is the exhaustive classification of typically 10^4-10^5 regions per image. Given that most of these regions contain uninformative background, the detector designs seem extremely superfluous and inefficient. In contrast, biological vision systems leverage selective attention for fast and efficient object detection. Recent neuroscientific findings shedding new light on the mechanism behind selective attention allowed us to formulate a new hypothesis of object detection efficiency and subsequently introduce a new object detection paradigm. To that end, we leverage this knowledge to design a novel region proposal network and empirically show that it achieves high object detection performance on the COCO dataset. Moreover, the model uses two to three orders of magnitude fewer computations than state-of-the-art models and consequently achieves inference speeds exceeding 500 frames/s, thereby making it possible to achieve object detection on embedded systems.
《FotonNet: A HW-Efficient Object Detection System Using 3D-Depth Segmentation and 2D-DNN Classifier》
arXiv：https://arxiv.org/abs/1811.07493
Object detection and classification is one of the most important computer vision problems. Ever since the introduction of deep learning \cite{krizhevsky2012imagenet}, we have witnessed a dramatic increase in the accuracy of this object detection problem. However, most of these improvements have occurred using conventional 2D image processing. Recently, low-cost 3D-image sensors, such as the Microsoft Kinect (Time-of-Flight) or the Apple FaceID (Structured-Light), can provide 3D-depth or point cloud data that can be added to a convolutional neural network, acting as an extra set of dimensions. In our proposed approach, we introduce a new 2D + 3D system that takes the 3D-data to determine the object region followed by any conventional 2D-DNN, such as AlexNet. In this method, our approach can easily dissociate the information collection from the Point Cloud and 2D-Image data and combine both operations later. Hence, our system can use any existing trained 2D network on a large image dataset, and does not require a large 3D-depth dataset for new training. Experimental object detection results across 30 images show an accuracy of 0.67, versus 0.54 and 0.51 for RCNN and YOLO, respectively.
《R2CNN++: Multi-Dimensional Attention Based Rotation Invariant Detector with Robust Anchor Strategy》
arXiv：https://arxiv.org/abs/1811.07126
Object detection plays a vital role in natural scene and aerial scene and is full of challenges. Although many advanced algorithms have succeeded in the natural scene, the progress in the aerial scene has been slow due to the complexity of the aerial image and the large degree of freedom of remote sensing objects in scale, orientation, and density. In this paper, a novel multi-category rotation detector is proposed, which can efficiently detect small objects, arbitrary direction objects, and dense objects in complex remote sensing images. Specifically, the proposed model adopts a targeted feature fusion strategy called inception fusion network, which fully considers factors such as feature fusion, anchor sampling, and receptive field to improve the ability to handle small objects. Then we combine the pixel attention network and the channel attention network to weaken the noise information and highlight the objects feature. Finally, the rotational object detection algorithm is realized by redefining the rotating bounding box. Experiments on public datasets including DOTA, NWPU VHR-10 demonstrate that the proposed algorithm significantly outperforms state-of-the-art methods. The code and models will be available at https://github.com/DetectionTeamUCAS/R2CNN-Plus-Plus_Tensorflow.
《Global and Local Sensitivity Guided Key Salient Object Re-augmentation for Video Saliency Detection》
arXiv：https://arxiv.org/abs/1811.07480
The existing still-static deep learning based saliency researches do not consider the weighting and highlighting of extracted features from different layers, all features contribute equally to the final saliency decision-making. Such methods always evenly detect all "potentially significant regions" and unable to highlight the key salient object, resulting in detection failure of dynamic scenes. In this paper, based on the fact that salient areas in videos are relatively small and concentrated, we propose a \textbf{key salient object re-augmentation method (KSORA) using top-down semantic knowledge and bottom-up feature guidance} to improve detection accuracy in video scenes. KSORA includes two sub-modules (WFE and KOS): WFE processes local salient feature selection using bottom-up strategy, while KOS ranks each object in global fashion by top-down statistical knowledge, and chooses the most critical object area for local enhancement. The proposed KSORA can not only strengthen the saliency value of the local key salient object but also ensure global saliency consistency. Results on three benchmark datasets suggest that our model has the capability of improving the detection accuracy on complex scenes. The significant performance of KSORA, with a speed of 17FPS on modern GPUs, has been verified by comparisons with other ten state-of-the-art algorithms.
《Pixel-Anchor: A Fast Oriented Scene Text Detector with Combined Networks》
arXiv：https://arxiv.org/abs/1811.07432
Recently, semantic segmentation and general object detection frameworks have been widely adopted by scene text detecting tasks. However, both of them alone have obvious shortcomings in practice. In this paper, we propose a novel end-to-end trainable deep neural network framework, named Pixel-Anchor, which combines semantic segmentation and SSD in one network by feature sharing and anchor-level attention mechanism to detect oriented scene text. To deal with scene text which has large variances in size and aspect ratio, we combine FPN and ASPP operation as our encoder-decoder structure in the semantic segmentation part, and propose a novel Adaptive Predictor Layer in the SSD. Pixel-Anchor detects scene text in a single network forward pass, no complex post-processing other than an efficient fusion Non-Maximum Suppression is involved. We have benchmarked the proposed Pixel-Anchor on the public datasets. Pixel-Anchor outperforms the competing methods in terms of text localization accuracy and run speed, more specifically, on the ICDAR 2015 dataset, the proposed algorithm achieves an F-score of 0.8768 at 10 FPS for 960 x 1728 resolution images.
《Improving Rotated Text Detection with Rotation Region Proposal Networks》
arXiv：https://arxiv.org/abs/1811.07031
A significant number of images shared on social media platforms such as Facebook and Instagram contain text in various forms. It's increasingly becoming commonplace for bad actors to share misinformation, hate speech or other kinds of harmful content as text overlaid on images on such platforms. A scene-text understanding system should hence be able to handle text in various orientations that the adversary might use. Moreover, such a system can be incorporated into screen readers used to aid the visually impaired. In this work, we extend the scene-text extraction system at Facebook, Rosetta, to efficiently handle text in various orientations. Specifically, we incorporate the Rotation Region Proposal Networks (RRPN) in our text extraction pipeline and offer practical suggestions for building and deploying a model for detecting and recognizing text in arbitrary orientations efficiently. Experimental results show a significant improvement on detecting rotated text.
《OrthoSeg: A Deep Multimodal Convolutional Neural Network for Semantic Segmentation of Orthoimagery》
arXiv：https://arxiv.org/abs/1811.07859
This paper addresses the task of semantic segmentation of orthoimagery using multimodal data e.g. optical RGB, infrared and digital surface model. We propose a deep convolutional neural network architecture termed OrthoSeg for semantic segmentation using multimodal, orthorectified and coregistered data. We also propose a training procedure for supervised training of OrthoSeg. The training procedure complements the inherent architectural characteristics of OrthoSeg for preventing complex co-adaptations of learned features, which may arise due to probable high dimensionality and spatial correlation in multimodal and/or multispectral coregistered data. OrthoSeg consists of parallel encoding networks for independent encoding of multimodal feature maps and a decoder designed for efficiently fusing independently encoded multimodal feature maps. A softmax layer at the end of the network uses the features generated by the decoder for pixel-wise classification. The decoder fuses feature maps from the parallel encoders locally as well as contextually at multiple scales to generate per-pixel feature maps for final pixel-wise classification resulting in segmented output. We experimentally show the merits of OrthoSeg by demonstrating state-of-the-art accuracy on the ISPRS Potsdam 2D Semantic Segmentation dataset. Adaptability is one of the key motivations behind OrthoSeg so that it serves as a useful architectural option for a wide range of problems involving the task of semantic segmentation of coregistered multimodal and/or multispectral imagery. Hence, OrthoSeg is designed to enable independent scaling of parallel encoder networks and decoder network to better match application requirements, such as the number of input channels, the effective field-of-view, and model capacity.
《M2U-Net: Effective and Efficient Retinal Vessel Segmentation for Resource-Constrained Environments》
arXiv：https://arxiv.org/abs/1811.07738
In this paper, we present a novel neural network architecture for retinal vessel segmentation that improves over the state of the art on two benchmark datasets, is the first to run in real time on high resolution images, and its small memory and processing requirements make it deployable in mobile and embedded systems. The M2U-Net has a new encoder-decoder architecture that is inspired by the U-Net. It adds pretrained components of MobileNetV2 in the encoder part and novel contractive bottleneck blocks in the decoder part that, combined with bilinear upsampling, drastically reduce the parameter count to 0.55M compared to 31.03M in the original U-Net. We have evaluated its performance against a wide body of previously published results on three public datasets. On two of them, the M2U-Net achieves new state-of-the-art performance by a considerable margin. When implemented on a GPU, our method is the first to achieve real-time inference speeds on high-resolution fundus images. We also implemented our proposed network on an ARM-based embedded system where it segments images in between 0.6 and 15 sec, depending on the resolution. Thus, the M2U-Net enables a number of applications of retinal vessel structure extraction, such as early diagnosis of eye diseases, retinal biometric authentication systems, and robot assisted microsurgery.
《Robust Visual Tracking using Multi-Frame Multi-Feature Joint Modeling》
arXiv：https://arxiv.org/abs/1811.07498
It remains a huge challenge to design effective and efficient trackers under complex scenarios, including occlusions, illumination changes and pose variations. To cope with this problem, a promising solution is to integrate the temporal consistency across consecutive frames and multiple feature cues in a unified model. Motivated by this idea, we propose a novel correlation filter-based tracker in this work, in which the temporal relatedness is reconciled under a multi-task learning framework and the multiple feature cues are modeled using a multi-view learning approach. We demonstrate the resulting regression model can be efficiently learned by exploiting the structure of blockwise diagonal matrix. A fast blockwise diagonal matrix inversion algorithm is developed thereafter for efficient online tracking. Meanwhile, we incorporate an adaptive scale estimation mechanism to strengthen the stability of scale variation tracking. We implement our tracker using two types of features and test it on two benchmark datasets. Experimental results demonstrate the superiority of our proposed approach when compared with other state-of-the-art trackers. project homepage：http://bmal.hust.edu.cn/project/KMF2JMTtracking.html
《Deep Siamese Networks with Bayesian non-Parametrics for Video Object Tracking》
arXiv：https://arxiv.org/abs/1811.07386
We present a novel algorithm utilizing a deep Siamese neural network as a general object similarity function in combination with a Bayesian optimization (BO) framework to encode spatio-temporal information for efficient object tracking in video. In particular, we treat the video tracking problem as a dynamic (i.e. temporally-evolving) optimization problem. Using Gaussian Process priors, we model a dynamic objective function representing the location of a tracked object in each frame. By exploiting temporal correlations, the proposed method queries the search space in a statistically principled and efficient way, offering several benefits over current state of the art video tracking methods.
《Exploit the Connectivity: Multi-Object Tracking with TrackletNet》
arXiv：https://arxiv.org/abs/1811.07258
Multi-object tracking (MOT) is an important and practical task related to both surveillance systems and moving camera applications, such as autonomous driving and robotic vision. However, due to unreliable detection, occlusion and fast camera motion, tracked targets can be easily lost, which makes MOT very challenging. Most recent works treat tracking as a re-identification (Re-ID) task, but how to combine appearance and temporal features is still not well addressed. In this paper, we propose an innovative and effective tracking method called TrackletNet Tracker (TNT) that combines temporal and appearance information together as a unified framework. First, we define a graph model which treats each tracklet as a vertex. The tracklets are generated by appearance similarity with CNN features and intersection-over-union (IOU) with epipolar constraints to compensate camera movement between adjacent frames. Then, for every pair of two tracklets, the similarity is measured by our designed multi-scale TrackletNet. Afterwards, the tracklets are clustered into groups which represent individual object IDs. Our proposed TNT has the ability to handle most of the challenges in MOT, and achieve promising results on MOT16 and MOT17 benchmark datasets compared with other state-of-the-art methods.
《Injecting and removing malignant features in mammography with CycleGAN: Investigation of an automated adversarial attack using neural networks》
arXiv：https://arxiv.org/abs/1811.07767
Purpose To train a cycle-consistent generative adversarial network (CycleGAN) on mammographic data to inject or remove features of malignancy, and to determine whether these AI-mediated attacks can be detected by radiologists. Material and Methods From the two publicly available datasets, BCDR and INbreast, we selected images from cancer patients and healthy controls. An internal dataset served as test data, withheld during training. We ran two experiments training CycleGAN on low and higher resolution images (256×256 px and 512×408 px). Three radiologists read the images and rated the likelihood of malignancy on a scale from 1-5 and the likelihood of the image being manipulated. The readout was evaluated by ROC analysis (Area under the ROC curve = AUC). Results At the lower resolution, only one radiologist exhibited markedly lower detection of cancer (AUC=0.85 vs 0.63, p=0.06), while the other two were unaffected (0.67 vs. 0.69 and 0.75 vs. 0.77, p=0.55). Only one radiologist could discriminate between original and modified images slightly better than guessing/chance (0.66, p=0.008). At the higher resolution, all radiologists showed significantly lower detection rate of cancer in the modified images (0.77-0.84 vs. 0.59-0.69, p=0.008), however, they were now able to reliably detect modified images due to better visibility of artifacts (0.92, 0.92 and 0.97). Conclusion A CycleGAN can implicitly learn malignant features and inject or remove them so that a substantial proportion of small mammographic images would consequently be misdiagnosed. At higher resolutions, however, the method is currently limited and has a clear trade-off between manipulation of images and introduction of artifacts.
《SEIGAN: Towards Compositional Image Generation by Simultaneously Learning to Segment, Enhance, and Inpaint》
arXiv：https://arxiv.org/abs/1811.07630
We present a novel approach to image manipulation and understanding by simultaneously learning to segment object masks, paste objects to another background image, and remove them from original images. For this purpose, we develop a novel generative model for compositional image generation, SEIGAN (Segment-Enhance-Inpaint Generative Adversarial Network), which learns these three operations together in an adversarial architecture with additional cycle consistency losses. To train, SEIGAN needs only bounding box supervision and does not require pairing or ground truth masks. SEIGAN produces better generated images (evaluated by human assessors) than other approaches and produces high-quality segmentation masks, improving over other adversarially trained approaches and getting closer to the results of fully supervised training.
《GAN-QP: A Novel GAN Framework without Gradient Vanishing and Lipschitz Constraint》
arXiv：https://arxiv.org/abs/1811.07296
We know SGAN may have a risk of gradient vanishing. A significant improvement is WGAN, with the help of 1-Lipschitz constraint on discriminator to prevent from gradient vanishing. Is there any GAN having no gradient vanishing and no 1-Lipschitz constraint on discriminator? We do find one, called GAN-QP. To construct a new framework of Generative Adversarial Network (GAN) usually includes three steps: 1. choose a probability divergence; 2. convert it into a dual form; 3. play a min-max game. In this articles, we demonstrate that the first step is not necessary. We can analyse the property of divergence and even construct new divergence in dual space directly. As a reward, we obtain a simpler alternative of WGAN: GAN-QP. We demonstrate that GAN-QP have a better performance than WGAN in theory and practice.
《Modeling Local Geometric Structure of 3D Point Clouds using Geo-CNN》
arXiv：https://arxiv.org/abs/1811.07782
Recent advances in deep convolutional neural networks (CNNs) have motivated researchers to adapt CNNs to directly model points in 3D point clouds. Modeling local structure has been proven to be important for the success of convolutional architectures, and researchers exploited the modeling of local point sets in the feature extraction hierarchy. However, limited attention has been paid to explicitly model the geometric structure amongst points in a local region. To address this problem, we propose Geo-CNN, which applies a generic convolution-like operation dubbed as GeoConv to each point and its local neighborhood. Local geometric relationships among points are captured when extracting edge features between the center and its neighboring points. We first decompose the edge feature extraction process onto three orthogonal bases, and then aggregate the extracted features based on the angles between the edge vector and the bases. This encourages the network to preserve the geometric structure in Euclidean space throughout the feature extraction hierarchy. GeoConv is a generic and efficient operation that can be easily integrated into 3D point cloud analysis pipelines for multiple applications. We evaluate Geo-CNN on ModelNet40 and KITTI and achieve state-of-the-art performance.
《PointConv: Deep Convolutional Networks on 3D Point Clouds》
arXiv：https://arxiv.org/abs/1811.07246
Unlike images which are represented in regular dense grids, 3D point clouds are irregular and unordered, hence applying convolution on them can be difficult. In this paper, we extend the dynamic filter to a new convolution operation, named PointConv. PointConv can be applied on point clouds to build deep convolutional networks. We treat convolution kernels as nonlinear functions of the local coordinates of 3D points comprised of weight and density functions. With respect to a given point, the weight functions are learned with multi-layer perceptron networks and the density functions through kernel density estimation. A novel reformulation is proposed for efficiently computing the weight functions, which allowed us to dramatically scale up the network and significantly improve its performance. The learned convolution kernel can be used to compute translation-invariant and permutation-invariant convolution on any point set in the 3D space. Besides, PointConv can also be used as deconvolution operators to propagate features from a subsampled point cloud back to its original resolution. Experiments on ModelNet40, ShapeNet, and ScanNet show that deep convolutional neural networks built on PointConv are able to achieve state-of-the-art on challenging semantic segmentation benchmarks on 3D point clouds. Besides, our experiments converting CIFAR-10 into a point cloud showed that networks built on PointConv can match the performance of convolutional networks in 2D images of a similar structure.
《Topology-Aware Non-Rigid Point Cloud Registration》
arXiv：https://arxiv.org/abs/1811.07014
In this paper, we introduce a non-rigid registration pipeline for pairs of unorganized point clouds that may be topologically different. Standard warp field estimation algorithms, even under robust, discontinuity-preserving regularization, tend to produce erratic motion estimates on boundaries associated with "close-to-open" topology changes. We overcome this limitation by exploiting backward motion: in the opposite motion direction, a "close-to-open' event becomes "open-to-close", which is by default handled correctly. At the core of our approach lies a general, topology-agnostic warp field estimation algorithm, similar to those employed in recently introduced dynamic reconstruction systems from RGB-D input. We improve motion estimation on boundaries associated with topology changes in an efficient post-processing phase. Based on both forward and (inverted) backward warp hypotheses, we explicitly detect regions of the deformed geometry that undergo topological changes by means of local deformation criteria and broadly classify them as "contacts" or `separations'. Subsequently, the two motion hypotheses are seamlessly blended on a local basis, according to the type and proximity of detected events. Our method achieves state-of-the-art motion estimation accuracy on the MPI Sintel dataset. Experiments on a custom dataset with topological event annotations demonstrate the effectiveness of our pipeline in estimating motion on event boundaries, as well as promising performance in explicit topological event detection.
《Past, Present, and Future Approaches Using Computer Vision for Animal Re-Identification from Camera Trap Data》
arXiv：https://arxiv.org/abs/1811.07749
The ability of a researcher to re-identify (re-ID) an individual animal upon re-encounter is fundamental for addressing a broad range of questions in the study of ecosystem function, community and population dynamics, and behavioural ecology. In this review, we describe a brief history of camera traps for re-ID, present a collection of computer vision feature engineering methodologies previously used for animal re-ID, provide an introduction to the underlying mechanisms of deep learning relevant to animal re-ID, highlight the success of deep learning methods for human re-ID, describe the few ecological studies currently utilizing deep learning for camera trap analyses, and our predictions for near future methodologies based on the rapid development of deep learning methods. By utilizing novel deep learning methods for object detection and similarity comparisons, ecologists can extract animals from an image/video data and train deep learning classifiers to re-ID animal individuals beyond the capabilities of a human observer. This methodology will allow ecologists with camera/video trap data to re-identify individuals that exit and re-enter the camera frame. Our expectation is that this is just the beginning of a major trend that could stand to revolutionize the analysis of camera trap data and, ultimately, our approach to animal ecology.
《CA3Net: Contextual-Attentional Attribute-Appearance Network for Person Re-Identification》
arXiv：https://arxiv.org/abs/1811.07544
Person re-identification aims to identify the same pedestrian across non-overlapping camera views. Deep learning techniques have been applied for person re-identification recently, towards learning representation of pedestrian appearance. This paper presents a novel Contextual-Attentional Attribute-Appearance Network (CA3Net) for person re-identification. The CA3Net simultaneously exploits the complementarity between semantic attributes and visual appearance, the semantic context among attributes, visual attention on attributes as well as spatial dependencies among body parts, leading to discriminative and robust pedestrian representation. Specifically, an attribute network within CA3Net is designed with an Attention-LSTM module. It concentrates the network on latent image regions related to each attribute as well as exploits the semantic context among attributes by a LSTM module. An appearance network is developed to learn appearance features from the full body, horizontal and vertical body parts of pedestrians with spatial dependencies among body parts. The CA3Net jointly learns the attribute and appearance features in a multi-task learning manner, generating comprehensive representation of pedestrians. Extensive experiments on two challenging benchmarks, i.e., Market-1501 and DukeMTMC-reID datasets, have demonstrated the effectiveness of the proposed approach.
《Re-Identification with Consistent Attentive Siamese Networks》
arXiv：https://arxiv.org/abs/1811.07487
We propose a new deep architecture for person re-identification (re-id). While re-id has seen much recent progress, spatial localization and view-invariant representation learning for robust cross-view matching remain key, unsolved problems. We address these questions by means of a new attention-driven Siamese learning architecture, called the Consistent Attentive Siamese Network. Our key innovations compared to existing, competing methods include (a) a flexible framework design that produces attention with only identity labels as supervision, (b) explicit mechanisms to enforce attention consistency among images of the same person, and (c) a new Siamese framework that integrates attention and attention consistency, producing principled supervisory signals as well as the first mechanism that can explain the reasoning behind the Siamese framework's predictions. We conduct extensive evaluations on the CUHK03-NP, DukeMTMC-ReID, and Market-1501 datasets, and establish a new state of the art, with our proposed method resulting in mAP performance improvements of 6.4%, 4.2%, and 1.2% respectively.
《Collaborative Dense SLAM》
arXiv：https://arxiv.org/abs/1811.07632
In this paper, we present a new system for live collaborative dense surface reconstruction. Cooperative robotics, multi participant augmented reality and human-robot interaction are all examples of situations where collaborative mapping can be leveraged for greater agent autonomy. Our system builds on ElasticFusion to allow a number of cameras starting with unknown initial relative positions to maintain local maps utilising the original algorithm. Carrying out visual place recognition across these local maps the system can identify when two maps overlap in space, providing an inter-map constraint from which the system can derive the relative poses of the two maps. Using these resulting pose constraints, our system performs map merging, allowing multiple cameras to fuse their measurements into a single shared reconstruction. The advantage of this approach is that it avoids replication of structures subsequent to loop closures, where multiple cameras traverse the same regions of the environment. Furthermore, it allows cameras to directly exploit and update regions of the environment previously mapped by other cameras within the system. We provide both quantitative and qualitative analyses using the synthetic ICL-NUIM dataset and the real-world Freiburg dataset including the impact of multi-camera mapping on surface reconstruction accuracy, camera pose estimation accuracy and overall processing time. We also include qualitative results in the form of sample reconstructions of room sized environments with up to 3 cameras undergoing intersecting and loopy trajectories.
《An Efficient Transfer Learning Technique by Using Final Fully-Connected Layer Output Features of Deep Networks》
arXiv：https://arxiv.org/abs/1811.07459
In this paper, we propose a computationally efficient transfer learning approach using the output vector of final fully-connected layer of deep convolutional neural networks for classification. Our proposed technique uses a single layer perceptron classifier designed with hyper-parameters to focus on improving computational efficiency without adversely affecting the performance of classification compared to the baseline technique. Our investigations show that our technique converges much faster than baseline yielding very competitive classification results. We execute thorough experiments to understand the impact of similarity between pre-trained and new classes, similarity among new classes, number of training samples in the performance of classification using transfer learning of the final fully-connected layer's output features.
《Transfer Learning with Deep CNNs for Gender Recognition and Age Estimation》
arXiv：https://arxiv.org/abs/1811.07344
In this project, competition-winning deep neural networks with pretrained weights are used for image-based gender recognition and age estimation. Transfer learning is explored using both VGG19 and VGGFace pretrained models by testing the effects of changes in various design schemes and training parameters in order to improve prediction accuracy. Training techniques such as input standardization, data augmentation, and label distribution age encoding are compared. Finally, a hierarchy of deep CNNs is tested that first classifies subjects by gender, and then uses separate male and female age models to predict age. A gender recognition accuracy of 98.7% and an MAE of 4.1 years is achieved. This paper shows that, with proper training techniques, good results can be obtained by retasking existing convolutional filters towards a new purpose.
《GLStyleNet: Higher Quality Style Transfer Combining Global and Local Pyramid Features》
arXiv：https://arxiv.org/abs/1811.07260
Recent studies using deep neural networks have shown remarkable success in style transfer especially for artistic and photo-realistic images. However, the approaches using global feature correlations fail to capture small, intricate textures and maintain correct texture scales of the artworks, and the approaches based on local patches are defective on global effect. In this paper, we present a novel feature pyramid fusion neural network, dubbed GLStyleNet, which sufficiently takes into consideration multi-scale and multi-level pyramid features by best aggregating layers across a VGG network, and performs style transfer hierarchically with multiple losses of different scales. Our proposed method retains high-frequency pixel information and low frequency construct information of images from two aspects: loss function constraint and feature fusion. Our approach is not only flexible to adjust the trade-off between content and style, but also controllable between global and local. Compared to state-of-the-art methods, our method can transfer not just large-scale, obvious style cues but also subtle, exquisite ones, and dramatically improves the quality of style transfer. We demonstrate the effectiveness of our approach on portrait style transfer, artistic style transfer, photo-realistic style transfer and Chinese ancient painting style transfer tasks. Experimental results indicate that our unified approach improves image style transfer quality over previous state-of-the-art methods, while also accelerating the whole process in a certain extent. Our code is available at https://github.com/EndyWon/GLStyleNet.
《Intention Oriented Image Captions with Guiding Objects》
arXiv：https://arxiv.org/abs/1811.07662
Although existing image caption models can produce promising results using recurrent neural networks (RNNs), it is difficult to guarantee that an object we care about is contained in generated descriptions, for example in the case that the object is inconspicuous in image. Problems become even harder when these objects did not appear in training stage. In this paper, we propose a novel approach for generating image captions with guiding objects (CGO). The CGO constrains the model to involve a human-concerned object, when the object is in the image, in the generated description while maintaining fluency. Instead of generating the sequence from left to right, we start description with a selected object and generate other parts of the sequence based on this object. To achieve this, we design a novel framework combining two LSTMs in opposite directions. We demonstrate the characteristics of our method on MSCOCO to generate descriptions for each detected object in images. With CGO, we can extend the ability of description to the objects being neglected in image caption labels and provide a set of more comprehensive and diverse descriptions for an image. CGO shows obvious advantages when applied to the task of describing novel objects. We show experiment results on both MSCOCO and ImageNet datasets. Evaluations show that our method outperforms the state-of-the-art models in the task with average F1 75.8, leading to better descriptions in terms of both content accuracy and fluency.
《Deep Comparison: Relation Columns for Few-Shot Learning》
arXiv：https://arxiv.org/abs/1811.07100
Few-shot deep learning is a topical challenge area for scaling visual recognition to open-ended growth in the space of categories to recognise. A promising line work towards realising this vision is deep networks that learn to match queries with stored training images. However, methods in this paradigm usually train a deep embedding followed by a single linear classifier. Our insight is that effective general-purpose matching requires discrimination with regards to features at multiple abstraction levels. We therefore propose a new framework termed Deep Comparison Network(DCN) that decomposes embedding learning into a sequence of modules, and pairs each with a relation module. The relation modules compute a non-linear metric to score the match using the corresponding embedding module's representation. To ensure that all embedding module's features are used, the relation modules are deeply supervised. Finally generalisation is further improved by a learned noise regulariser. The resulting network achieves state of the art performance on both miniImageNet and tieredImageNet, while retaining the appealing simplicity and efficiency of deep metric learning approaches.
《iQIYI-VID: A Large Dataset for Multi-modal Person Identification》
arXiv：https://arxiv.org/abs/1811.07548
Person identification in the wild is very challenging due to great variation in poses, face quality, clothes, makeup and so on. Traditional research, such as face recognition, person re-identification, and speaker recognition, often focuses on a single modal of information, which is inadequate to handle all the situations in practice. Multi-modal person identification is a more promising way that we can jointly utilize face, head, body, audio features, and so on. In this paper, we introduce iQIYI-VID, the largest video dataset for multi-modal person identification. It is composed of 600K video clips of 5,000 celebrities. These video clips are extracted from 400K hours of online videos of various types, ranging from movies, variety shows, TV series, to news broadcasting. All video clips pass through a careful human annotation process, and the error rate of labels is lower than 0.2%. We evaluated the state-of-art models of face recognition, person re-identification, and speaker recognition on the iQIYI-VID dataset. Experimental results show that these models are still far from being perfect for task of person identification in the wild. We further demonstrate that a simple fusion of multi-modal features can improve person identification considerably. We have released the dataset online to promote multi-modal person identification research.
《Addressing the Invisible: Street Address Generation for Developing Countries with Deep Learning》
NIPS 2018 Workshop
arXiv：https://arxiv.org/abs/1811.07769
More than half of the world's roads lack adequate street addressing systems. Lack of addresses is even more visible in daily lives of people in developing countries. We would like to object to the assumption that having an address is a luxury, by proposing a generative address design that maps the world in accordance with streets. The addressing scheme is designed considering several traditional street addressing methodologies employed in the urban development scenarios around the world. Our algorithm applies deep learning to extract roads from satellite images, converts the road pixel confidences into a road network, partitions the road network to find neighborhoods, and labels the regions, roads, and address units using graph- and proximity-based algorithms. We present our results on a sample US city, and several developing cities, compare travel times of users using current ad hoc and new complete addresses, and contrast our addressing solution to current industrial and open geocoding alternatives.
《Handwriting Recognition of Historical Documents with few labeled data》
arXiv：https://arxiv.org/abs/1811.07768
Historical documents present many challenges for offline handwriting recognition systems, among them, the segmentation and labeling steps. Carefully annotated textlines are needed to train an HTR system. In some scenarios, transcripts are only available at the paragraph level with no text-line information. In this work, we demonstrate how to train an HTR system with few labeled data. Specifically, we train a deep convolutional recurrent neural network (CRNN) system on only 10% of manually labeled text-line data from a dataset and propose an incremental training procedure that covers the rest of the data. Performance is further increased by augmenting the training set with specially crafted multiscale data. We also propose a model-based normalization scheme which considers the variability in the writing scale at the recognition phase. We apply this approach to the publicly available READ dataset. Our system achieved the second best result during the ICDAR2017 competition.
《GroundNet: Segmentation-Aware Monocular Ground Plane Estimation with Geometric Consistency》
arXiv：https://arxiv.org/abs/1811.07222
We focus on the problem of estimating the orientation of the ground plane with respect to a mobile monocular camera platform (e.g., ground robot, wearable camera, assistive robotic platform). To address this problem, we formulate the ground plane estimation problem as an inter-mingled multi-task prediction problem by jointly optimizing for point-wise surface normal direction, 2D ground segmentation, and depth estimates. Our proposed model -- GroundNet -- estimates the ground normal in two streams separately and then a consistency loss is applied on top of the two streams to enforce geometric consistency. A semantic segmentation stream is used to isolate the ground regions and are used to selectively back-propagate parameter updates only through the ground regions in the image. Our experiments on KITTI and ApolloScape datasets verify that the GroundNet is able to predict consistent depth and normal within the ground region. It also achieves top performance on ground plane normal estimation and horizon line detection.
《Image-to-GPS Verification Through A Bottom-Up Pattern Matching Network》
arXiv：https://arxiv.org/abs/1811.07288
The image-to-GPS verification problem asks whether a given image is taken at a claimed GPS location. In this paper, we treat it as an image verification problem -- whether a query image is taken at the same place as a reference image retrieved at the claimed GPS location. We make three major contributions: 1) we propose a novel custom bottom-up pattern matching (BUPM) deep neural network solution; 2) we demonstrate that the verification can be directly done by cross-checking a perspective-looking query image and a panorama reference image, and 3) we collect and clean a dataset of 30K pairs query and reference. Our experimental results show that the proposed BUPM solution outperforms the state-of-the-art solutions in terms of both verification and localization.
《Matching RGB Images to CAD Models for Object Pose Estimation》
arXiv：https://arxiv.org/abs/1811.07249
We propose a novel method for 3D object pose estimation in RGB images, which does not require pose annotations of objects in images in the training stage. We tackle the pose estimation problem by learning how to establish correspondences between RGB images and rendered depth images of CAD models. During training, our approach only requires textureless CAD models and aligned RGB-D frames of a subset of object instances, without explicitly requiring pose annotations for the RGB images. We employ a deep quadruplet convolutional neural network for joint learning of suitable keypoints and their associated descriptors in pairs of rendered depth images which can be matched across modalities with aligned RGB-D views. During testing, keypoints are extracted from a query RGB image and matched to keypoints extracted from rendered depth images, followed by establishing 2D-3D correspondences. The object's pose is then estimated using the RANSAC and PnP algorithms. We conduct experiments on the recently introduced Pix3D dataset and demonstrate the efficacy of our proposed approach in object pose estimation as well as generalization to object instances not seen during training.
《Optical Flow Dataset and Benchmark for Visual Crowd Analysis》
arXiv：https://arxiv.org/abs/1811.07170
The performance of optical flow algorithms greatly depends on the specifics of the content and the application for which it is used. Existing and well established optical flow datasets are limited to rather particular contents from which none is close to crowd behavior analysis; whereas such applications heavily utilize optical flow. We introduce a new optical flow dataset exploiting the possibilities of a recent video engine to generate sequences with ground-truth optical flow for large crowds in different scenarios. We break with the development of the last decade of introducing ever increasing displacements to pose new difficulties. Instead we focus on real-world surveillance scenarios where numerous small, partly independent, non rigidly moving objects observed over a long temporal range pose a challenge. By evaluating different optical flow algorithms, we find that results of established datasets can not be transferred to these new challenges. In exhaustive experiments we are able to provide new insight into optical flow for crowd analysis. Finally, the results have been validated on the real-world UCF crowd tracking benchmark while achieving competitive results compared to more sophisticated state-of-the-art crowd tracking approaches.
《Simulating LIDAR Point Cloud for Autonomous Driving using Real-world Scenes and Traffic Flows》
arXiv：https://arxiv.org/abs/1811.07112
We present a LIDAR simulation framework that can automatically generate 3D point cloud based on LIDAR type and placement. The point cloud, annotated with ground truth semantic labels, is to be used as training data to improve environmental perception capabilities for autonomous driving vehicles. Different from previous simulators, we generate the point cloud based on real environment and real traffic flow. More specifically we employ a mobile LIDAR scanner with cameras to capture real world scenes. The input to our simulation framework includes dense 3D point cloud and registered color images. Moving objects (such as cars, pedestrians, bicyclists) are automatically identified and recorded. These objects are then removed from the input point cloud to restore a static background (e.g., environment without movable objects). With that we can insert synthetic models of various obstacles, such as vehicles and pedestrians in the static background to create various traffic scenes. A novel LIDAR renderer takes the composite scene to generate new realistic LIDAR points that are already annotated at point level for synthetic objects. Experimental results show that our system is able to close the performance gap between simulation and real data to be 1 ~ 6% in different applications, and for model fine tuning, only 10% ~ 20% extra real data could help to outperform the original model trained with full real dataset.
《DSCnet: Replicating Lidar Point Clouds with Deep Sensor Cloning》
arXiv：https://arxiv.org/abs/1811.07070
Convolutional neural networks (CNNs) have become increasingly popular for solving a variety of computer vision tasks, ranging from image classification to image segmentation. Recently, autonomous vehicles have created a demand for depth information, which is often obtained using hardware sensors such as Light detection and ranging (LIDAR). Although it can provide precise distance measurements, most LIDARs are still far too expensive to sell in mass-produced consumer vehicles, which has motivated methods to generate depth information from commodity automotive sensors like cameras. In this paper, we propose an approach called Deep Sensor Cloning (DSC). The idea is to use Convolutional Neural Networks in conjunction with inexpensive sensors to replicate the 3D point-clouds that are created by expensive LIDARs. To accomplish this, we develop a new dataset (DSDepth) and a new family of CNN architectures (DSCnets). While previous tasks such as KITTI depth prediction use an interpolated RGB-D images as ground-truth for training, we instead use DSCnets to directly predict LIDAR point-clouds. When we compare the output of our models to a $75,000 LIDAR, we find that our most accurate DSCnet achieves a relative error of 5.77% using a single camera and 4.69% using stereo cameras.