Search

This paper tackles automated screening of Retinopathy of Prematurity (ROP), one of the most common causes of visual loss in childhood. Clinically, ROP screening per case requires multiple color fundus image instances that capture different zones of the (premature) retina. A desirable model shall not only make a decision at the case level, but also pinpoint which instances and what part of the instances are responsible for the decision. This paper makes the first attempt to accomplish three tasks, i.e. ROP case classification, instance selection and abnormality localization in a unified framework. To that end, we propose a new model that effectively combines instance-attention based deep multiple instance learning (MIL) and spatial attention (SA). The propose model, which we term MIL-SA, identifies positive instances in light of their contributions to case-level decision. Meanwhile, abnormal regions in the identified instances are automatically localized by the SA mechanism. Moreover, MIL-SA is learned from case-level binary labels exclusively, and in an end-to-end manner. Experiments on a large clinical dataset of 2,186 cases with 11,053 fundus images show the viability of the proposed model for all the three tasks.

During the process of space travelling and space landing, the spacecraft attitude estimation is the indispensable work for navigation. Since there are not enough satellites for GPS-like localization in space, the computer vision technique is adopted to address the issue. The most crucial task for localization is the extraction of correspondences. In computer vision, optical flow estimation is often used for finding correspondences between images. As the deep neural network being more popular in recent years, FlowNet2 has played a vital role which achieves great success. In this paper, we present PA-FlowNet, an end-to-end pose-auxiliary optical flow network which can use the predicted relative camera pose to improve the performance of optical flow. PA-FlowNet is composed of two sub-networks, the foreground-attention flow network and the pose regression network. The foreground-attention flow network is constructed by FlowNet2 model and modified with the proposed foreground-attention approach. We introduced this approach with the concept of curriculum learning for foreground-background segmentation to avoid backgrounds from resulting in flow prediction error. The pose regression network is used to regress the relative camera pose as an auxiliary for increasing the accuracy of the flow estimation. In addition, to simulate the test environment for spacecraft pose estimation, we construct a 64K moon model and simulate aerial photography with various attitudes to generate Moon64K dataset in this paper. PA-FlowNet significantly outperforms all existing methods on the proposed Moon64K dataset. Furthermore, we also predict the relative camera pose via proposed PA-FlowNet and accomplish the remarkable performance.

The notion of anchor plays a major role in modern detection algorithms such as the Faster-RCNN [1] or the SSD detector [2]. Anchors relate the features of the last layers of the detector with bounding boxes containing objects in images. Despite their importance, the literature on object detection has not paid real attention to them. The motivation of this paper comes from the observations that (i) each anchor learns to classify and regress candidate objects independently (ii) insufficient examples are available for each anchor in case of small-scale datasets. This paper addresses these questions by proposing a novel hierarchical head for the SSD detector. The new design has the added advantage of no extra weights, as compared to the original design at inference time, while improving detectors performance for small size training sets. Improved performance on PASCAL-VOC and state-of-the-art performance on FlickrLogos-47 validate the method. We also show when the proposed design does not give additional performance gain over the original design.

The key to an accurate understanding of terrain is to extract the informative features from the multi-modal data obtained from different devices. Sensors, such as RGB cameras, depth sensors, vibration sensors, and microphones, are used as the multi-modal data. Many studies have explored ways to use them, especially in the robotics field. Some papers have successfully introduced single-modal or multi-modal methods. However, in practice, robots can be faced with extreme conditions; microphones do not work well in crowded scenes, and an RGB camera cannot capture terrains well in the dark. In this paper, we present a novel framework using the multimodal variational autoencoder and the Gaussian mixture model clustering algorithm on image data and audio data for terrain type clustering. Our method enables the terrain type clustering even if one of the modalities (either image or audio) is missing at the test-time. We evaluated the clustering accuracy with a conventional multi-modal terrain type clustering method and we conducted ablation studies to show the effectiveness of our approach.

Monocular depth estimation plays a key role in 3D scene understanding, and a number of recent papers have achieved significant improvements using deep learning based algorithms. Most papers among them proposed methods that use a pre-trained network as a deep feature extractor and then decode the obtained features to create a depth map. In this study, we focus on how to use this encoder-decoder structure to deliver meaningful representation throughout the entire network. We propose a new network architecture with our suggested modules to create a more accurate depth map by bridging the context between the encoding and decoding phase. First, we place the pyramid block at the bottleneck of the network to enlarge the view and convey rich information about the global context to the decoder. Second, we suggest a skip connection with the fuse module to aggregate the encoder and decoder feature. Finally, we validate our approach on the NYU Depth V2 and KITTI datasets. The experimental results show the efficacy of the suggested model and show performance gains over the state-of-the-art model.

The prevalent adoptions of GPS-enabled devices have witnessed an explosion of various location-based services which produces a huge amount of trajectories monitoring the individuals' movements. In this paper, we tackle Trajectory-User Link (TUL) problem, which identifies humans' movement patterns and links trajectories to the users who generated them. Existing solutions on TUL problem employ recurrent neural networks and variational autoencoder methods, which face the bottlenecks in the case of excessively long trajectories and fragmentary users' movements. However, these are common characteristics of trajectory data in reality, leading to performance degradation of the existing models. In this paper, we propose an end-to-end attention recurrent neural learning framework, called TULAR (Trajectory-User Link with Attention Recurrent Networks), which focus on selected parts of the source trajectories when linking. TULAR introduce the Trajectory Semantic Vector (TSV) via unsupervised location representation learning and recurrent neural networks, by which to reckon the weight of parts of source trajectory. Further, we employ three attention scores for the weight measurements. Experiments are conducted on two real world datasets and compared with several existing methods, and the results show that TULAR yields a new state-of-the-art performance. Source code is public available at GitHub: https://github.com/taos123/TULAR.

This paper tackles the problem of human action recognition, defined as classifying which action is displayed in a trimmed sequence, from skeletal data. Albeit state-of-the-art approaches designed for this application are all supervised, in this paper we pursue a more challenging direction: solving the problem with unsupervised learning. To this end, we propose a novel subspace clustering method, which exploits covariance matrix to enhance the action's discriminability and a timestamp pruning approach that allow us to better handle the temporal dimension of the data. Through a broad experimental validation, we show that our computational pipeline surpasses existing unsupervised approaches but also can result in favorable performances as compared to the supervised methods. The code is available here: https://github.com/IIT-PAVIS/subspace-clustering-action-recognition

In a society with increasing age, the understanding of human falls it is of paramount importance. This paper presents a Decision Support System whose pipeline is designed to extract and compute physical domain's features achieving the state of the art accuracy on the Le2i and UR fall detection datasets. The paper uses the Kinematic Theory of Rapid Human Movement and its sigma-lognormal model together with classic physical features to achieve 98% and 99% of accuracy in automatic fall detection on respectively Le2i and URFD datasets. The effort made in the design of this work is toward recognition of falls by using physical models whose laws are clear and understandable.

Group retrieval has attracted plenty of attention in artificial intelligence, traditional group retrieval researches assume that members in a group are unique and do not change under different cameras. However, the assumption may not be met for practical situations such as open-world and group-ambiguity scenarios. This paper tackles an important yet non-studied problem: re-identifying changing groups of people under the open-world and group-ambiguity scenarios in different camera fields. The open-world scenario considers that there are probably non-target people for the probe set appear in the searching gallery, while the group-ambiguity scenario means the group members may change. The open-world and group-ambiguity issue is very challenging for the existing methods because the changing of group members results in dramatic visual variations. Nevertheless, as far as we know, the existing literature lacks benchmarks which target on coping with this issue. In this paper, we propose a new group retrieval dataset named OWGA-Campus to consider these challenges. Moreover, we propose a person-to-group similarity matching based ambiguity removal (P2GSM-AR) method to solve these problems and realize the intention of group retrieval. Experimental results on OWGA-Campus dataset demonstrate the effectiveness and robustness of the proposed P2GSM-AR approach in improving the performance of the state-of-the-art feature extraction methods of person re-id towards the open-world and ambiguous group retrieval task.

This paper aims to develop a deep neural network for Continuous Sign Language Recognition (CSLR) with iterative Gloss Recognition (GR) fine-tuning. CSLR has been a popular research field in the last few years and iterative optimization methods are well established. This paper introduces our proposed architecture involving Spatiotemporal feature-extraction model to segment useful “gloss-unit” features and BiLSTM with CTC as a sequence model. Spatiotemporal Feature Extractor is used for both image features extraction and sequence length reduction. To this end, we compare different architectures for feature extraction and sequence model. In addition, we iteratively fine-tune feature extractor on gloss-unit video segments with alignments from the end2end model. During the iterative training, we use novel alignment correction technique, which is based on minimum transformations of Levenshtein distance. All the experiments are conducted on the RWTH-PHOENIX-Weather-2014 dataset.

Hand gesture and motion sensing offer an intuitive and natural form of human-machine interface. Air-writing systems allow users to draw alpha-numerical or linguistic characters in the virtual board in air through hand gestures. Traditionally, radar-based air-writing systems have been based on a network of radars, at least three, to localize the hand target through trilateration algorithm followed by tracking to extract the drawn trajectory, which is then followed by recognition of the drawn character by either Long-Short Term Memory (LSTM) utilizing the sensed trajectory or Deep Convolutional Neural Network (DCNN) utilizing a reconstructed 2D image from the trajectory. However, the practical deployments of such systems are limited since the detection of the finger or hand target by all three radars cannot be guaranteed leading to failure of the trilateration algorithm. Further placement of three or more radars for the air-writing solution is neither always physically plausible nor cost-effective. Furthermore, these solutions do not exploit the full potentials of deep neural networks, which are generally capable of learning features implicitly. In this paper, we propose an air-writing system based on a network of sparse radars, i.e. strictly less than three, using 1D DCNN-LSTM-1D transposed DCNN architecture to reconstruct and classify the drawn character utilizing only the range information from each radar. The paper employs real data using one and two 60 GHz milli-meter wave radar sensors to demonstrate the success of the proposed air-writing solution.

This paper studies the indoor localisation of WiFi devices based on a commodity chipset and standard channel sounding. First, we present a novel shallow neural network (SNN) in which features are extracted from the channel state information (CSI) corresponding to WiFi subcarriers received on different antennas and used to train the model. The single-layer architecture of this localisation neural network makes it lightweight and easy-to-deploy on devices with stringent constraints on computational resources. We further investigate for localisation the use of deep learning models and design novel architectures for convolutional neural network (CNN) and long-short term memory (LSTM). We extensively evaluate these localisation algorithms for continuous tracking in indoor environments. Experimental results prove that even an SNN model, after a careful handcrafted feature extraction, can achieve accurate localisation. Meanwhile, using a well-organised architecture, the neural network models can be trained directly with raw data from the CSI and localisation features can be automatically extracted to achieve accurate position estimates. We also found that the performance of neural network-based methods are directly affected by the number of anchor access points (APs) regardless of their structure. With three APs, all neural network models proposed in this paper can obtain localisation accuracy of around 0.5 metres. In addition the proposed deep NN architecture reduces the data pre-processing time by 6.5 hours compared with a shallow NN using the data collected in our testbed. In the deployment phase, the inference time is also significantly reduced to 0.1 ms per sample. We also demonstrate the generalisation capability of the proposed method by evaluating models using different target movement characteristics to the ones in which they were trained., Comment: Accepted for presentation at the 25th International Conference on Pattern Recognition (ICPR), IEEE, 2020

Snapshot hyperspectral imaging can capture the 3D hyperspectral image (HSI) with a single 2D measurement and has attracted increasing attention recently. Recovering the underlying HSI from the compressive measurement is an ill-posed problem and exploiting the image prior is essential for solving this ill-posed problem. However, existing reconstruction methods always start from modeling image prior with the 1D vector or 2D matrix and cannot fully exploit the structurally spectral-spatial nature in 3D HSI, thus leading to a poor fidelity. In this paper, we propose an effective high-order tensor optimization based method to boost the reconstruction fidelity for snapshot hyperspectral imaging. We first build high-order tensors by exploiting the spatial-spectral correlation in HSI. Then, we propose a weight high-order singular value regularization (WHOSVR) based low-rank tensor recovery model to characterize the structure prior of HSI. By integrating the structure prior in WHOSVR with the system imaging process, we develop an optimization framework for HSI reconstruction, which is finally solved via the alternating minimization algorithm. Extensive experiments implemented on two representative systems demonstrate that our method outperforms state-of-the-art methods., This paper has been accepted by IEEE ICPR2020

As reported by Hong Kong Government in 2017, there are more than 1.5 million residents suffering from hearing impairment in Hong Kong. Most of them rely on Hong Kong Sign Language for daily communication while there are only 63 registered sign language interpreters in Hong Kong. To address this specific social issue and also facilitate the effective communication between the hearing impaired and other people, this paper introduces a word-level Hong Kong Sign Language(HKSL) dataset which currently includes 45 isolated words and at least 30 sign videos per word performed by different signers(more than 1500 videos in total now and still enlarging). Based on this dataset, this paper systemically compares the performances of various deep learning approaches, including (1) 2D histogram of oriented gradients(HOG) feature/pose estimation/feature extraction with long-short term memory(LSTM) layer; (2) 3D Residual Neural Network(ResNet) (3) (2+1)D Residual Neural Network, in HKSL recognition. Meanwhile, to further improve the accuracy of sign language recognition, this paper proposes a novel method called (3+2+1)D ResNet Model with Frame Selection which adopts blurriness detection with Laplacian kernel to construct high-quality video clips and also combines both (2+1)D and 3D ResNet for recognizing the sign language. At the end, the experimental results show that the proposed method outperforms other deep learning approaches and attains an impressive accuracy of 94.6% in our dataset.

In this paper we present a novel defogging technique,named CurL-Defog, aimed at minimizing the creation of unwanted artifacts during the defogging process. The majority of learning based defogging approaches rely on paired data (i.e.,the same images with and without fog), where fog is artificially added to clear images: this often provides good results on mildly fogged images but does not generalize well to real difficult cases. On the other hand, the models trained with real unpaired data (e.g. CycleGAN) can provide visually impressive results but they often produce unwanted artifacts. In this paper we propose a curriculum learning strategy coupled with an enhanced CycleGAN model in order to reduce the number of produced artifacts, while maintaining state-of-the-art performance in terms of contrast enhancement and image reconstruction. We also introduce a new metric, called HArD (Hazy Artifact Detector) to numerically quantify the amount of artifacts in the defogged images, thus avoiding the tedious and subjective manual inspection of the results. The proposed approach compares favorably with state-of-the-art techniques on both real and synthetic datasets.

This paper relies on the Mean Decision Rule (MDR) method for solving large-scale binary SVM problems. It consists in taking small random samples of the full dataset and separate training for each of them with consecutive averaging the respective individual decision rules to obtain a final one. This paper proposes two new approaches to improve it. The first proposed approach is a new sampling technique that exploits SVM and MDR properties to fast form so-called smart samples by selecting only the objects, that are candidates to be the support ones. The proposed technique essentially increases MDR convergence and allows to reach the highest quality in less time. In the case of kernel-based MDR (KMDR) the proposed sampling technique allows additionally to reduce the number of support objects in the final decision rule and, as a result, to decrease the recognition time. The second proposed approach is a new data strategy to accelerate random access to large datasets stored in the traditional libsvm format. The proposed strategy allows to quickly extract random subsets of objects from a file and load them into RAM and is also suitable for any sampling-based methods, including stochastic gradient methods. Joint use of the proposed approaches with (K)MDR allows to obtain the best (or near the best) decision of large-scale binary SVM problems faster, compared to the existing SVM solvers.

This paper focuses on two fundamental graph recognition tasks: node classification and graph classification. Existing methods usually learn the node and graph representations for these two tasks separately, and ignore modeling the relations between the local and global structures. In this paper, we propose a general approach to learn the local and global features collaboratively: (1) in order to characterize the correlation among nodes and communities (a set of nodes), we employ the joint community detection/dynamic routing modules to generate the clustering assignment matrices at first and then utilize these matrices to cluster nodes to capture the global information of graphs (locally relevant graph representations). Inspired by the success of spectral clustering, we minimize the ratiocut loss to help optimize the learned assignment matrices. (2) We maximize the mutual information between local and global representations to help learn the globally relevant node representations. Experimental results on a variety of node and graph classification benchmarks show that our model can achieve superior performance over the state-of-the-art approaches.

Deep Convolutional Neural Networks (DCNN), despite imitating the human visual system, present no such decision credibility as human observers. This phenomenon, therefore, leads to the limitations of DCNN's applications in the security and trusted computing, such as self-driving cars and medical diagnosis. Focusing on this issue, our work aims to explore the way DCNN makes decisions. In this paper, the major contributions we made are: firstly, provide the hypothesis, “point-wise activation” of convolution function, according to the analysis of DCNN's architectures and training process; secondly, point out the effect of “point-wise activation” on DCNN's uninterpretable classification and pool robustness, and then suggest, in particular, the contradiction between the traditional and DCNN's convolution kernel functions; finally, distinguish decision-making interpretability from semantic interpretability, and indicate that DCNN's decision-making mechanism need to evolve towards the direction of semantics in the future. Besides, the “point-wise activation” hypothesis and conclusions proposed in our paper are supported by extensive experimental results.

Photo retouching features are being integrated into a growing number of mobile applications. Current learning-based approaches enhance images using large convolutional neural network-based models, where the result is received directly from the neural network outputs. This method can lead to artifacts in the resulting images, models that are complicated to interpret, and can be computationally expensive. In this paper, we explore the application of a filter-based approach in order to overcome the problems outlined above. We focus on creating a lightweight solution suitable for use on mobile devices when designing our model. A significant performance increase was achieved through implementing consistency regularization used in semi-supervised learning. The proposed model can be used on mobile devices and achieves competitive results compared to known models. In this paper, we explore the application of a filter-based approach in order to overcome the problems outlined above. We focus on creating a lightweight solution suitable for use on mobile devices when designing our model. A significant performance increase was achieved through implementing consistency regularization used in semi-supervised learning. The proposed model can be used on mobile devices and achieves competitive results compared to known models.

AC machines may be subjected to different electrical or mechanical faults during their operation. Fault patterns can be detected in the DC current from the machine's E-Drive system with the help of Deep or Machine Learning algorithms. However, Deep or Machine Learning algorithms require large amounts of dataset for training and without the availability of a large dataset the algorithms fail to generalize or give their optimal performance. Collecting large amounts of data from faulty machine can be a tedious task. It is expensive and not always possible. In some cases, the machine is completely damaged even before sufficient amount of data can be collected. Also, data collection from defected machine may cause permanent damage to the connected system. Therefore, in this paper the problem of small dataset is tackled by presenting a methodology for large dataset generation by using the well-known generative model, Generative Adversarial Networks (GAN). As an example, the stator open circuit fault in a synchronous machine is considered. DC currents from the machine's E-Drive system are measured from different healthy and faulty machines and are used for training of two 1d DCGANs (Deep Convolutional GANs), one for the healthy and the other for the current signal from the faulty machine. Conventional GANs are difficult to train, however in this paper, training parameters of 1d DCGAN are tuned which results an improved training process. The performance of generator during the training of 1d DCGAN is evaluated by using the Frechet Inception Distance (FID) metric. The proposed 1d DCGAN model is said to converge when FID score between the real and generated signal reaches below a certain threshold. The generated signals from the trained 1d DCGAN are further evaluated using the PDF (Probability Density Function), frequency domain analysis and other measures which check for duplication of the real data and their statistical diversity. The trained 1d DCGAN is able to generate DC current signals for building large datasets for the training of Deep or Machine learning models.

In this paper we propose a novel method for location recognition and prediction in smart homes based on semi-supervised learning. We use data collected from low-resolution depth video cameras installed in four apartments with older adults over 70 years of age, and collected during a period of one to seven weeks. The location of the person in the depth images is detected by a person detection algorithm adapted from YOLO (You Only Look Once). The locations extracted from the videos are then clustered using K-means clustering. Sequence prediction algorithms are used to predict the next cluster (location) based on the previous clusters (locations). The accuracy of predicting the next location is up to 91%, a significant improvement compared to the case where binary sensors are placed in the apartment based on human intuition. The paper presents an analysis on the effect of the memory length (i.e. the number of previous clusters used to predict the next one), and on the amount of recorded data required to converge.

This paper studies the problem of Person Re-Identification (ReID) for large-scale applications. Recent research efforts have been devoted to building complicated part models, which introduce considerably high computational cost and memory consumption, inhibiting its practicability in large-scale applications. This paper aims to develop a novel learning strategy to find efficient feature embeddings while maintaining the balance of accuracy and model complexity. More specifically, we find by enhancing the classical triplet loss together with cross-entropy loss, our method can explore the hard examples and build a discriminant feature embedding yet compact enough for large-scale applications. Our method is carried out progressively using Bayesian optimization, and we call it the Progressive Learning Algorithm (PLA). Extensive experiments on three large-scale datasets show that our PLA is comparable or better than the-state-of-the-arts. Especially, on the challenging Market-ISOI dataset, we achieve Rank-1=94.7%/mAP=89.4% while saving at least 30 % parameters than strong part models.

Facial micro-expressions are spontaneous, subtle, involuntary muscle movements occurring briefly on the face. The spotting and recognition of these expressions are difficult due to the subtle behavior, and the time duration of these expressions is about half a second, which makes it difficult for humans to identify them. These micro-expressions have many applications in our daily life, such as in the field of online learning, game playing, lie detection, and therapy sessions. Traditionally, researchers use RGB images/videos to spot and classify these micro-expressions, which pose challenging problems, such as illumination, privacy concerns and pose variation. The use of depth videos solves these issues to some extent, as the depth videos are not susceptible to the variation in illumination. This paper describes the collection of a first RGB-D dataset for the classification of facial micro-expressions into 6 universal expressions: Anger, Happy, Sad, Fear, Disgust, and Surprise. This paper shows the comparison between the RGB and Depth videos for the classification of facial micro-expressions. Further, a comparison of results shows that depth videos alone can be used to classify facial micro-expressions correctly in a decision tree structure by using the traditional and deep learning approaches with good classification accuracy. The dataset will be released to the public in the near future.

In decades many gait recognition methods have been proposed using different techniques. However, due to a real-world scenario of clothing variations, a reduction of the recognition rate occurs for most of these methods. Thus in this paper, a part-based spatio-temporal feature learning method is proposed to tackle the problem of clothing variations for gait recognition. First, based on the anatomical properties, human bodies are segmented into two regions, which are affected and unaffected by clothing variations. A learning network is particularly proposed in this paper to grasp principal spatio-temporal features from those unaffected regions. Different from most part-based methods with spatial or temporal features solely being utilized, in our method these two features are associated in a more collaborative manner. Snapshots are created for each gait sequence from the $H-W$ and $T-W$ views. Stable spatial information is embedded in the $H-W$ view and adequate temporal information is embedded in the $T-W$ view. An inherent relationship exists between these two views. Thus, a collaborative spatio-temporal feature will be hybridized by concatenating these correlative spatial and temporal information. The robustness and efficiency of our proposed method are validated by experiments on CASIA Gait Dataset B and OU-ISIR Treadmill Gait Dataset B. Our proposed method can both achieve the state-of-the-art results on these two databases.

Feature-based transfer is one of the most effective methodologies for transfer learning. Existing studies usually assume that the learned new feature representation is truly domain-invariant, and thus directly train a transfer model $M$ on source domain. In this paper, we consider a more realistic scenario where the new feature representation is suboptimal and small divergence still exists across domains. We propose a new learning strategy with a transfer model called Randomized Transferable Machine (RTM). More specifically, we work on source data with the new feature representation learned from existing feature-based transfer methods. The key idea is to enlarge source training data populations by randomly corrupting source data using some noises, and then train a transfer model $\widetilde{\mathcal{M}}$ that performs well on all the corrupted source data populations. In principle, the more corruptions are made, the higher the probability of the target data can be covered by the constructed source populations. and thus better transfer performance can be achieved by $\widetilde{\mathcal{M}}$ An ideal case is with infinite corruptions, which however is infeasible in reality. We develop a marginalized solution with linear regression model and dropout noise. With a marginalization trick, we can train an RTM that is equivalently to training using infinite source noisy populations without truly conducting any corruption. More importantly, such an RTM has a closed-form solution, which enables very fast and efficient training. Extensive experiments on various real-world transfer tasks show that RTM is a promising transfer model.

This paper proposes a simple yet effective strategy for improving deep supervised learning, named Stochastic Label Refinery (SLR), by refining training labels to more informative labels. When training a neural network, target distributions (or ground-truth) are typically “hard”, which means the target label of each category consists of only 0 and 1. However, the fixed “hard” target distributions do not capture association between categories or that between objects. In this study, instead of using the hard target distributions, we iteratively generate “soft” target label distributions for training the neural networks, which leads to better performances. The soft target distributions are obtained via an Expectation-Maximization (EM) iteration, where the “true” target distributions and the learned models are regarded as hidden variables. In E step, the models are optimized to approximate the target distributions on stochastic splits of training data; In M step, the target distributions are updated with predicted pseudo-label on leave-out splits. Extensive experiments on classification and ordinal regression tasks, empirically prove that the refined target distribution consistently leads to considerable performance improvements even applied on competitive baselines. Notably, in DeepDR 2020 Diabetic Retinopathy Grading (DeepDRiD) challenge, our method improves the quadratic weighted kappa on official validation set from 0.8247 to 0.8348 and achieves a state-of-the-art score on online test set. The proposed SLR technique is easy to implement and practically applicable.

In low light condition, the captured color (RGB) images are highly degraded by noise with severe texture loss. In this paper, we propose deep fusion of RGB and NIR paired images in low light condition using convolutional neural networks (CNNs). The proposed deep fusion network consists of three independent sub-networks: DenoisingNet, EnhancingNet, and FusionNet. We build a denoising sub-network to eliminate noise from noisy RGB images. After denoising, we perform an enhancing sub-network to increase the brightness of low light RGB images. Since NIR image contains fine details, we fuse it with the Y channel of RGB image through a fusion subnetwork. Experimental results demonstrate that the proposed method successfully fuses RGB and NIR images, and generates high quality fusion results containing textures and colors.

Recently, many Near infrared-visible (NIR-VIS) heterogeneous face recognition (HFR) methods have been proposed in the community. But it remains a challenging problem because of the sensing gap along with large pose variations. In this paper, we propose an Exemplar Guided Cross-Spectral Face Hallucination (EGCH) to reduce the domain discrepancy through disentangled representation learning. For each modality, EGCH contains a spectral encoder as well as a structure encoder to disentangle spectral and structure representation, respectively. It also contains a traditional generator that reconstructs the input from the above two representations, and a structure generator that predicts the facial parsing map from the structure representation. Besides, mutual information minimization and maximization are conducted to boost disentanglement and make representations adequately expressed. Then the translation is built on structure representations between two modalities. Provided with the transformed NIR structure representation and original VIS spectral representation, EGCH is capable to produce high-fidelity VIS images that preserve the topology structure of the input NIR while transfer the spectral information of an arbitrary VIS exemplar. Extensive experiments demonstrate that the proposed method achieves more promising results both qualitatively and quantitatively than the state-of-the-art NIR-VIS methods.

In this paper we propose a method for Offline Handwritten Mathematical Expression recognition. The method is a fast and accurate thanks to its architecture, which include both a Convolutional Neural Network and a Recurrent Neural Network. The CNN extracts features from the image to recognize and its output is provided to the RNN which produces the mathematical expression encoded in the LATEX language. To process both sequential and non-sequential mathematical expressions we also included a deconvolutional module which, in a recursive way, segments the image for additional analysis trough a recursive process. The results obtained show a very high accuracy obtained on a large handwritten data set of 9100 samples of handwritten expressions.

Subspace clustering finds a multi-subspace representation that best fits a high-dimensional dataset. The computational and storage complexities of existing algorithms limit their usefulness for large scale data. In this paper, we develop a novel scalable approach to subspace clustering termed Fast Direction-Search-Based Subspace Clustering (Fast DiSC). In sharp contrast to existing scalable solutions which are mostly based on the self-expressiveness property of the data, Fast DiSC rests upon a new representation obtained from projections on computed data-dependent directions. These directions are derived from a convex formulation for optimal direction search to gauge hidden similarity relations. The computational complexity is significantly reduced by performing direction search in partitions of sampled data, followed by a retrieval step to cluster out-of-sample data using projections on the computed directions. A theoretical analysis underscores the ability of the proposed formulation to construct local similarity relations for the different data points. Experiments on both synthetic and real data demonstrate that the proposed algorithm can often outperform the state-of-the-art clustering methods.

Video-based person re-identification has received increasing attention recently, as it plays an important role within surveillance video analysis. Video-based Re-Id is an expansion of earlier image-based re-identification methods by learning features from a video via multiple image frames for each person. Most contemporary video Re-ID methods utilise complex CNN-based network architectures using 3D convolution or multi-branch networks to extract spatial-temporal video features. By contrast, in this paper, we illustrate superior performance from a simple single stream 2D convolution network leveraging the ResNet50-IBN architecture to extract frame-level features followed by temporal attention for clip level features. These clip level features can be generalised to extract video level features by averaging without any significant additional cost. Our approach uses best video Re-Id practice and transfer learning between datasets to outperform existing state-of-the-art approaches on the MARS, PRID2011 and iLIDS-VID datasets with 89.62%, 97.75%, 97.33% rank-1 accuracy respectively and with 84.61% mAP for MARS, without reliance on complex and memory intensive 3D convolutions or multi-stream networks architectures as found in other contemporary work. Conversely, our work shows that global features extracted by the 2D convolution network are a sufficient representation for robust state of the art video Re-ID.

Sketch colorization is usually expensive and time-consuming for artists, and automating this process can have many pragmatic applications in the animation, comic book, and video game industry. However, automatic image colorization faces many challenges, because sketches not only lack texture information but also potentially entail complicated objects that require acute coloring. These difficulties usually result in incorrect color assignments that can ruin the aesthetic appeal of the final output. In this paper, we present a novel component-based matching framework that combines deep learned features and quadratic programming with a new cost function to solve this colorization problem. The proposed framework inputs a character's sketches as well as a colored image in the same cut of a movie, and outputs a high-quality sequence of colorized frames based on the color assignment in the reference colored image. To carry out this colorization task, we first utilize a pretrained ResNet-34 model to extract elementary components' features to match certain pairs of components (one component from the sketch and one from reference). Next, a graph representation is constructed in order to process and match the remaining components that could not be done in the first step. Since the first step has reduced the number of components to be matched by the graph, we can solve this graph problem in a short computing time even when there are hundreds of different components present in each sketch. We demonstrate the effectiveness of the proposed solution by conducting comprehensive experiments and producing aesthetically pleasing results. To the best of our knowledge, our framework is the first work that combines deep learning extraction and graph representation to colorize anime sketches and achieves a high pixel-level accuracy at a reasonable time cost.

Melanoma is the deadliest form of skin cancer and it is diagnosed mainly visually, starting from initial clinical screening and followed by dermoscopic analysis, biopsy and histopathological examination. A dermatologist's recognition of melanoma may be subject to errors and may take some time to diagnose it. In this regard, deep learning can be useful in the study and classification of skin cancer. In particular, by classifying images with Deep Neural Network methodologies, it is possible to obtain comparable or even superior results compared to those of dermatologists. In this paper, we propose a methodology for the classification of melanoma by adopting different deep learning techniques applied to a common dataset, composed of images from the ISIC dataset and consisting of different types of skin diseases, including melanoma on which we applied a specific pre-processing phase. In particular, a comparison of the results is performed in order to select the best effective neural network to be applied to the problem of recognition and classification of melanoma. Moreover, we also evaluate the impact of the preprocessing phase on the final classification. Different metrics such as accuracy, sensitivity, and specificity have been selected to assess the goodness of the adopted neural networks and compare them also with the manual classification of dermatologists.

Activation in deep neural networks is fundamental to achieving non-linear mappings. Traditional studies mainly focus on finding fixed activations for a particular set of learning tasks or model architectures. The research on flexible activation is quite limited in both designing philosophy and application scenarios. In this study, three principles of choosing flexible activation components are proposed and a general combined form of flexible activation functions is implemented. Based on this, a novel family of flexible activation functions that can replace sigmoid or tanh in LSTM cells are implemented, as well as a new family by combining ReLU and ELUs. Also, two new regularisation terms based on assumptions as prior knowledge are introduced. It has been shown that LSTM models with proposed flexible activations P-Sig- Ramp provide significant improvements in time series forecasting, while the proposed P-E2-ReLU achieves better and more stable performance on lossy image compression tasks with convolutional auto-encoders. In addition, the proposed regularization terms improve the convergence, performance and stability of the models with flexible activation functions. The code for this paper is available at https://github.com/9NXJRDDRQK/Flexible_Activation.

The ubiquity of smart speakers is increasing, with a growing number of households utilising these devices to make calls over the telephony network. As the technology is typically configured to retain the cellular phone number of the user, it presents challenges in applications where knowledge of the true call origin is required. There are a wide range of makes and models for these devices, as is the case with cell phones, and it is challenging to detect the general category as a smart speaker or cell, independent of the designated phone number. In this paper, we present an approach to differentiate between calls originating from smart speakers and ones from cellular devices using only the audio. We present a feature set that characterises the relevant acoustic information, such as the degree of reverberation and noise, to distinguish between these categories. When evaluated on a dataset spanning multiple models for each device category, as well as different modes-of-usage and microphone-speaker distances, the method yields an Equal Error Rate (EER) of 12.6%.

In this paper, we propose a method for human pose estimation from a Low-resolution Far-InfraRed (LFIR) image sequence captured by a 16 × 16 FIR sensor array. Human body estimation from such a single LFIR image is a hard task. For training the estimation model, annotation of the human pose to the images is also a difficult task for human. Thus, we propose the LFIR2Pose model which accepts a sequence of LFIR images and outputs the human pose of the last frame, and also propose an automatic annotation system for the model training. Additionally, considering that the scale of human body motion is largely different among body parts, we also propose a loss function focusing on the difference. Through an experiment, we evaluated the human pose estimation accuracy with an original data set, and confirmed that human pose can be estimated accurately from an LFIR image sequence.

Scene flow estimation based on stereo sequences is a comprehensive task relevant to disparity and optical flow. Some existing methods are time-consuming and often fail in the presence of reflective surfaces. In this paper, we propose a two-stage adaptive object scene flow estimation method using a hybrid CNN-CRF model (ACOSF), which benefits from high-quality features and the structured modelling capability. Meanwhile, in order to balance the computational efficiency and accuracy, we employ adaptive iteration for energy function optimization, which is flexible and efficient for various scenes. Besides, we utilize high-quality pixel selection to reduce the computation time with only a slight decrease in accuracy. Our method achieves competitive results with the state-of-the-art, which ranks second on the challenging KITTI 2015 scene flow benchmark.

International audience; The way people consume sports on TV has drastically evolved in the last years, particularly under the combined effects of the legalization of sport betting and the huge increase of sport analytics. Several companies are nowadays sending observers in the stadiums to collect live data of all the events happening on the field during the match. Those data contain meaningful information providing a very detailed description of all the actions occurring during the match to feed the coaches and staff, the fans, the viewers, and the gamblers. Exploiting all these data, sport broadcasters want to generate extra content such as match highlights, match summaries, players and teams analytics, etc., to appeal subscribers. This paper explores the problem of summarizing professional soccer matches as automatically as possible using both the aforementioned event-stream data collected from the field and the content broadcasted on TV. We have designed an architecture, introducing first (1) a Multiple Instance Learning method that takes into account the sequential dependency among events and then (2) a hierarchical multimodal attention layer that grasps the importance of each event in an action. We evaluate our approach on matches from two professional European soccer leagues, showing its capability to identify the best actions for automatic summarization by comparing with real summaries made by human operators.

As the specialty of hyperspectral image (HSI), it consists of 2D spatial and 1D spectral information. In the field of deep learning, HSI classification is an appealing research topic. Many existing methods process the HSI in spatial or spectral domain separately, which cannot fully extract the representative features, and the most used 3D convolutional neural network (3D-CNN) will suffer from mixing up complex spectral information. In this paper, we propose a spatial-spectral unified method by using recurrent neural networks (RNN) and multi-scanning direction strategy to construct spatial-spectral information sequences for learning the spatial dependencies among the central pixel and neighboring pixels. Meanwhile, residual connections and dense connections are introduced into multi-scanning direction sequences to overcome the memory problem in the RNN. The proposed method got 99.58% and 99.81% accuracy respectively on two benchmark datasets: the Pavia University dataset and the Pavia Center dataset. It demonstrates the proposed method can achieve state-of-the-art results.

3D face reconstruction from a single 2D facial image is a challenging and concerned problem. Recent methods based on CNN typically aim to learn parameters of 3D Morphable Model (3DMM) from 2D images to render face alignment and 3D face reconstruction. Most algorithms are designed for faces with small, medium yaw angles, which is extremely challenging to align faces in large poses. At the same time, they are not efficient usually. The main challenge is that it takes time to determine the parameters accurately. In order to address this challenge with the goal of improving performance, this paper proposes a novel and efficient end-to-end framework. We design an efficient and lightweight network model combined with Depthwise Separable Convolution and Muti-scale Representation, Lightweight Attention Mechanism, named Mobile-FRNet. Simultaneously, different loss functions are used to constrain and optimize 3DMM parameters and 3D vertices during training to improve the performance of the network. Meanwhile, extensive experiments on the challenging datasets show that our method significantly improves the accuracy of face alignment and 3D face reconstruction. Model parameters and complexity of our method are also improved greatly.

Ranking is a ubiquitous problem setting that appears in many real-world applications. The superior players or objects in the ranking list are oftentimes estimated from matchups and pairwise comparisons. In order to predict the result of unobserved events, various models have been developed to integrate the matchup and pairwise comparison results into a single ranking list of players. Amongst these models, the Bradley-Terry model is the mainstream which achieves the goal by authorizing enriched probabilistic interpretation of players. However, the model suffers from its strong assumption of transitive relationships and becomes vulnerable in practice when intransitive relationships exist. In this paper, we propose a low-rank matrix approach to characterize all players and unify the related works on representation learning of players by rearranging the constructively. Our experimental results on synthetic datasets and real-world datasets show that the proposed model is competitive with the standard models in terms of out-of-sample predictive performance and consistency of model interpretation.

In this paper, we propose a multi-attribute regression network (MARN) to investigate the problem of face reconstruction, especially in challenging cases when faces undergo large variations including severe poses, extreme expressions, and partial occlusions in unconstrained environments. The traditional 3DMM parametric regression method does not distinguish the learning of identity, expression, and attitude attributes, resulting in lacking geometric details in the reconstructed face. We propose to learn a face multi-attribute features during 3D face reconstruction from single 2D images. Our MARN enables the network to better extract the feature information of face identity, expression, and pose attributes. We introduce three loss functions to constrain the above three face attributes respectively. At the same time, we carefully design the geometric contour constraint loss function, using the constraints of sparse 2D face landmarks to improve the reconstructed geometric contour information. The experimental results show that our MARN has achieved significant improvements in 3D face reconstruction and face alignment on the AFLW2000-3D and AFLW datasets.

In this paper, we look into the problem of disentangled representation learning and controllable image synthesis in a deep generative model. We develop an encoder-decoder architecture for a variant of the Variational Auto-Encoder (VAE) with two latent codes $z_{1}$ and $z_{2}$ . Our framework uses $z_{2}$ to capture specified factors of variation while $z_{1}$ captures the complementary factors of variation. To this end, we analyze the learning problem from the perspective of multivariate mutual information, derive optimizable lower bounds of the conditional mutual information in the image synthesis processes and incorporate them into the training objective. We validate our method empirically on the Color MNIST dataset and the CelebA dataset by showing controllable image syntheses. Our proposed paradigm is simple yet effective and is applicable to many situations, including those where there is not an explicit factorization of features available, or where the features are non-categorical.

The paper presents a new theory of combined moment invariants to Gaussian blur and spatial affine transformation. The blur kernel may be arbitrary oriented, scaled and elongated. No prior information about the kernel parameters and about the underlaying affine transform is required. The main idea, expressed by the Substitution Theorem, is to substitute pure blur invariants into traditional affine moment invariants. Potential applications of the new descriptors are in blur-invariant image recognition and in robust template matching.

Conference Name: 2020 25th International Conference on Pattern Recognition (ICPR) Date of Conference: 10-15 January 2021 There is an increasing demand on utilizing camera equipped drones and their applications in many domains varying from agriculture to entertainment and from sports events to surveillance. In such drone applications, an essential and a common task is tracking an object of interest visually. Drone (or UAV) images have different properties when compared to the ground taken (natural) images and those differences introduce additional complexities to the existing object trackers to be directly applied on drone applications. Some important differences among those complexities include (i) smaller object sizes to be tracked and (ii) different orientations and viewing angles yielding different texture and features to be observed. Therefore, new algorithms trained on drone images are needed for the drone-based applications. In this paper, we introduce a deep reinforcement learning (RL) based single object tracker that tracks an object of interest in drone images by estimating a series of actions to find the location of the object in the next frame. This is the first work introducing a single object tracker using a deep RL-based technique for drone images. Our proposed solution introduces a novel reward function that aims to reduce the total number of actions taken to estimate the object's location in the next frame and also introduces a different backbone network to be used on low resolution images. Additionally, we introduce a set of new actions into the action library to better deal with the above-mentioned complexities. We compare our proposed solutions to a state of the art tracking algorithm from the recent literature and demonstrate up to 3.87 % improvement in precision and 3.6% improvement in IoU values on the VisDrone2019 data set. We also provide additional results on OTB-100 data set and show up to 3.15% improvement in precision on the OTB-100 data set when compared to the same previous state of the art algorithm. Lastly, we analyze the ability to handle some of the challenges faced during tracking, including but not limited to occlusion, deformation, and scale variation for our proposed solutions.

Contextualized representations from a pre-trained language model are central to achieve a high performance on downstream NLP task. The pre-trained BERT and A Lite BERT (ALBERT) models can be fine-tuned to give state-ofthe-art results in sentence-pair regressions such as semantic textual similarity (STS) and natural language inference (NLI). Although BERT-based models yield the [CLS] token vector as a reasonable sentence embedding, the search for an optimal sentence embedding scheme remains an active research area in computational linguistics. This paper explores on sentence embedding models for BERT and ALBERT. In particular, we take a modified BERT network with siamese and triplet network structures called Sentence-BERT (SBERT) and replace BERT with ALBERT to create Sentence-ALBERT (SALBERT). We also experiment with an outer CNN sentence-embedding network for SBERT and SALBERT. We evaluate performances of all sentence-embedding models considered using the STS and NLI datasets. The empirical results indicate that our CNN architecture improves ALBERT models substantially more than BERT models for STS benchmark. Despite significantly fewer model parameters, ALBERT sentence embedding is highly competitive to BERT in downstream NLP evaluations., 6 pages, 2 figures, to be published in 25th International Conference on Pattern Recognition, ICPR2020

This paper presents a new approach to the problem of detecting out-of-distribution (OOD) inputs in image classifications with deep convolutional networks. We leverage so-called boundary samples to enforce low confidence (maximum softmax probabilities) for inputs far away from the training data. In particular, we propose the boundary optimised samples (named BoS) training algorithm for generating them. Unlike existing approaches, it does not require extra generative adversarial network, but achieves the goal by simply back propagating the gradient of an appropriately designed loss function to the input samples. At the end of the BoS training, all the boundary samples are in principle located on a specific level hypersurface with respect to the designed loss. Our contributions are i) the BoS training as an efficient alternative to generate boundary samples, ii) a robust algorithm therewith to enforce low confidence for OOD samples, and iii) experiments demonstrating improved OOD detection over the baseline. We show the performance using standard datasets for training and different test sets including Fashion MNIST, EMNIST, SVHN, and CIFAR-100, preceded by evaluations with a synthetic 2-dimensional dataset that provide an insight for the new procedure.

In recent years, learning based shadow detection and removal approaches have shown prospects and, in most cases, yielded state-of-the-art results. The performance of these approaches, however, relies heavily on the construction of training database of shadow images, shadow-free versions, and shadow maps as the ground truth. This conventional data gathering is time-consuming, expensive, or even practically intractable to realize especially for outdoor scenes with complicated shadow patterns, thus limiting the size of the data available for training. In this paper, we leverage on large high quality synthetic image database and domain adaptation to mitigate the bottlenecks resulting from insufficient training samples and domain bias. Specifically, our approach utilizes adversarial training to predict near-pixel-perfect shadow map from synthetic shadow image for downstreaming shadow removal steps. At inference time, we capitalize on domain adaptation via image style transfer to map the style of real-world scene to that of synthetic scene for the purpose of detecting and subsequently removing shadow. Comprehensive experiments indicate that our approach outperforms state-of-the-art methods on selected benchmark datasets.

The Softmax prediction function is widely used to train Deep Convolutional Neural Networks (DCNNs) for large-scale face recognition and other applications. The limitation of the softmax activation is that the resulting probability distribution always has a full support. This full support leads to larger intraclass variations. In this paper, we formulate a novel loss function, called Angular Sparsemax for face recognition. The proposed loss function promotes sparseness of the hypotheses prediction function similar to Sparsemax [1] with Fenchel-Young regularisation. By introducing an additive angular margin on the score vector, the discriminatory power of the face embedding is further improved. The proposed loss function is experimentally validated on several databases in terms of recognition accuracy. Its performance compares well with the state of the art Arcface loss.

Wrist-wearable devices such as smartwatch hardware have gained popularity as they provide quick access to various information and easy access to multiple applications. Among the numerous smartwatch applications, user verification based on the handwriting is gaining momentum by considering its reliability and user-friendliness. In this paper, we present a novel technique for user verification using a smartwatch based writing pattern or style. The proposed approach leverages accelerometer data captured from the smartwatch that is further represented using 2D Continuous Wavelet Transform (CWT) and deep features extracted using the pre-trained ResNet50. These features are classified using an ensemble of classifiers to make the final decision on user verification. Extensive experiments are carried out on a newly captured dataset using two different smartwatches with three different writing scenarios (or activities). Experimental results provide critical insights and analysis of the results in such a verification scenario.