Search

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.

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

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

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 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.

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.

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.

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.

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

Uncertainty estimates help to identify ambiguous, novel, or anomalous inputs, but the reliable quantification of uncertainty has proven to be challenging for modern deep networks. In order to improve uncertainty estimation, we propose On-Manifold Adversarial Data Augmentation or OMADA, which specifically attempts to generate the most challenging examples by following an on-manifold adversarial attack path in the latent space of an autoencoder-based generative model that closely approximates decision boundaries between two or more classes. On a variety of datasets as well as on multiple diverse network architectures, OMADA consistently yields more accurate and better calibrated classifiers than baseline models, and outperforms competing approaches such as Mixup, as well as achieving similar performance to (at times better than) post-processing calibration methods such as temperature scaling. Variants of OMADA can employ different sampling schemes for ambiguous on-manifold examples based on the entropy of their estimated soft labels, which exhibit specific strengths for generalization, calibration of predicted uncertainty, or detection of out-of-distribution inputs., Accepted for oral at International Conference on Pattern Recognition, ICPR 2020. Nominated (top 4) for Best Industry Related Paper Award (BIRPA)

General graphs are difficult for learning due to their irregular structures. Existing works employ message passing along graph edges to extract local patterns using customized graph kernels, but few of them are effective for the integration of such local patterns into global features. In contrast, in this paper we study the methods to transfer the graphs into trees so that explicit orders are learned to direct the feature integration from local to global. To this end, we apply the breadth first search (BFS) to construct trees from the graphs, which adds direction to the graph edges from the center node to the peripheral nodes. In addition, we proposed a novel projection scheme that transfer the trees to image representations, which is suitable for conventional convolution neural networks (CNNs) and recurrent neural networks (RNNs). To best learn the patterns from the graph-tree-images, we propose TreeRNN, a 2D RNN architecture that recurrently integrates the image pixels by rows and columns to help classify the graph categories. We evaluate the proposed method on several graph classification datasets, and manage to demonstrate comparable accuracy with the state-of-the-art on MUTAG, PTC-MR and NCI1 datasets.

Weakly supervised object localization is a challenging task in which the object of interest should be localized while learning its appearance. State-of-the-art methods recycle the architecture of a standard CNN by using the activation maps of the last layer for localizing the object. While this approach is simple and works relatively well, object localization relies on different features than classification, thus, a specialized localization mechanism is required during training to improve performance. In this paper, we propose a convolutional, multi-scale spatial localization network that provides accurate localization for the object of interest. Experimental results on CUB-200-2011 and ImageNet datasets show that our proposed approach provides competitive performance for weakly supervised localization., Comment: 8 pages, 9 figures

International audience; Several recent studies have proposed an automatic plant disease identification system based on deep learning. However, these approaches are generally based on learned classification models with target classes of joint host species-disease pairs that may not allow optimal use of the available information. This is because these approaches require distinguishing between similar host species or diseases, and more importantly, they have limited scalability due to the size of a network gradually increases as new classes are added, despite the fact that information on host species or diseases is already available. This constraint is all the more important as it can be difficult to collect/establish a specific list of all diseases for each host plant species in an actual application. In this paper, we address the problems by proposing a new conditional multi-task learning (CMTL) approach which allows the distribution of host species and disease characteristics learned simultaneously with a conditional link between them. This conditioning is formed in such a way that the knowledge to infer the prediction of one concept (the diseases) depends on the other concept (the host species), which corresponds to the way plant pathologists used to infer the diseases of the host species. We show that our approach can improve the performance of plant disease identification compared to the usual species-disease pair modeling in the previous studies. Meanwhile, we also compose a new dataset on plant disease identification that could serve as an important benchmark in this field

Recognizing human activities from multi-channel time series data collected from wearable sensors has become an important practical application of machine learning. A serious challenge comes from the presence of coherent activities or body movements, such as movements of the head while walking or sitting, since signals representing these movements are mixed and interfere with each other. Basic multi-label classification typically assumes independence within the multiple activities. This is oversimplified and reduces modeling power even when using state-of-the-art deep learning methods. In this paper, we investigate this new problem, which we name “Coherent Human Activity Recognition (Co-HAR)”, that keeps complete conditional dependency between the multiple labels. Additionally, we treat Co-HAR as a dense labelling problem that classifies each sample on a time step with multiple coherent labels to provide high-fidelity and duration-sensitive support to high-precision applications. To explicitly model conditional dependency, a novel condition-aware deep architecture “Conditional-UNet” is developed to allow for multiple dense labeling for Co-HAR. We also contribute a first-of-its-kind Co-HAR dataset for head gesture recognition associated with a user's activity, walking or sitting, to the research community. Extensive experiments on this dataset show that our model outperforms state-of-the-art deep learning methods and achieves up to 92% accuracy on context-based head gesture classification.

Lung segmentation on CT images is a crucial step for a computer-aided diagnosis system of lung diseases. The existing deep learning based lung segmentation methods are less efficient to segment lungs on clinical CT images, especially that the segmentation on lung boundaries is not accurate enough due to complex pulmonary opacities in practical clinics. In this paper, we propose a boundary-guided network (BG-Net) to address this problem. It contains two auxiliary branches that seperately segment lungs and extract the lung boundaries, and an aggregation branch that efficiently exploits lung boundary cues to guide the network for more accurate lung segmentation on clinical CT images. We evaluate the proposed method on a private dataset collected from the Osaka university hospital and four public datasets including StructSeg [1], HUG [2], VESSEL12 [3], and a Novel Coronavirus 2019 (COVID-19) dataset [4]. Experimental results show that the proposed method can segment lungs more accurately and outperform several other deep learning based methods.

With the prosperous development of Convolutional Neural Networks, currently they can perform excellently on visual understanding tasks when the input images are high quality and common quality images. However, large degradation in performance always occur when the input images are low quality images. In this paper, we propose a new super-resolution method in order to improve the classification performance for low-resolution images. In an image, the regions in which pixel values vary dramatically contain more abundant high frequency contents compared to other parts. Based on this fact, we design a weight map and integrate it with a super-resolution CNN training framework. During the process of training, this weight map can find out positions of the high frequency pixels in ground truth high-resolution images. After that, the pixel-level loss function takes effect only at these found positions to minimize the difference between reconstructed high-resolution images and ground truth high-resolution images. Compared with other state-of-the-art super-resolution methods, the experiment results show that our method can recover more high frequency contents in high-resolution image reconstructing, and better improve the classification accuracy after low-resolution image preprocessing.

Given an unknown audio source, the estimation of time differences-of-arrivals (TDOAs) can be efficiently and robustly solved using blind channel identification and exploiting the cross-correlation identity (CCI). Prior “blind” works have improved the estimate of TDOAs by means of different algorithmic solutions and optimization strategies, while always sticking to the case $N=2$ microphones. But what if we can obtain a direct improvement in performance by just increasing $N$ ? In this paper we try to investigate this direction, showing that, despite the arguable simplicity, this is capable of (sharply) improving upon state-of-the-art blind channel identification methods based on CCI, without modifying the computational pipeline. Inspired by our results, we seek to warm up the community and the practitioners by paving the way (with two concrete, yet preliminary, examples) towards joint approaches in which advances in the optimization are combined with an increased number of microphones, in order to achieve further improvements.

Identifying vehicles across a network of cameras with non-overlapping fields of view remains a challenging research problem due to scene occlusions, significant inter-class similarity and intra-class variability. In this paper, we propose an end-to-end multi-level re-identification network that is capable of successfully projecting same identity vehicles closer to one another in the embedding space, compared to vehicles of different identities. Robust feature representations are obtained by combining features at multiple levels of the network. As for the learning process, we employ a recent state-of-the-art structured metric learning loss function previously applied to other retrieval problems and adjust it to the vehicle re-identification task. Furthermore, we explore the cases of image-to-image, image-to-video and video-to-video similarity metric. Finally, we evaluate our system and achieve great performance on two large-scale publicly available datasets, CityFlow-ReID and VeRi-776. Compared to most existing state-of-art approaches, our approach is simpler and more straightforward, utilizing only identity-level annotations, while avoiding post-processing the ranking results (re-ranking) at the testing phase.

In recent years, surveillance cameras are widely deployed in public places, and the general crime rate has been reduced significantly due to these ubiquitous devices. Usually, these cameras provide cues and evidence after crimes are conducted, while they are rarely used to prevent or stop criminal activities in time. It is both time and labor consuming to manually monitor a large amount of video data from surveillance cameras. Therefore, automatically recognizing violent behaviors from video signals becomes essential. This paper summarizes several existing video datasets for violence detection and proposes the RWF-2000 database with 2,000 videos captured by surveillance cameras in real-world scenes. Also, we present a new method that utilizes both the merits of 3D-CNNs and optical flow, namely Flow Gated Network. The proposed approach obtains an accuracy of 87.25% on the test set of our proposed database. The database and source codes are currently open to access., Accepted by ICPR'20

Facial makeup style transfer is an extremely challenging sub-field of image-to-image-translation. Due to this difficulty, state-of-the-art results are mostly reliant on the Face Parsing Algorithm, which segments a face into parts in order to easily extract makeup features. However, this algorithm can only work well on high-definition images where facial features can be accurately extracted. Faces in many real-world photos, such as those including a large background or multiple people, are typically of low-resolution, which considerably hinders state-of-the-art algorithms. In this paper, we propose an end-to-end holistic approach to effectively transfer makeup styles between two low-resolution images. The idea is built upon a novel weighted multi-scale spatial attention module, which identifies salient pixel regions on low-resolution images in multiple scales, and uses channel attention to determine the most effective attention map. This design provides two benefits: low-resolution images are usually blurry to different extents, so a multi-scale architecture can select the most effective convolution kernel size to implement spatial attention; makeup is applied on both a macro-level (foundation, fake tan) and a micro-level (eyeliner, lipstick) so different scales can excel in extracting different makeup features. We develop an Augmented CycleGAN network that embeds our attention modules at selected layers to most effectively transfer makeup. Our system is tested with the FBD data set, which consists of many low-resolution facial images, and demonstrate that it outperforms state-of-the-art methods, particularly in transferring makeup for blurry images and partially occluded images.

Comparison of tree based representations for image sequences; International audience; Morphological hierarchies form a popular framework aiming at emphasizing the multiscale structure of digital image by performing an unsupervised spatial partitioning of the data. These hierarchies have been recently extended to cope with image sequences, and different strategies have been proposed to allow their construction from spatio-temporal data. In this paper, we compare these hierarchical representation strategies for image sequences according to their structural properties. We introduce a projection method to make these representations comparable. Furthermore, we extend one of these recent strategies in order to obtain more efficient hierarchical representations for image sequences. Experiments were conducted on both synthetic and real datasets, the latter being made of satellite image time series. We show that building one hierarchy by using spatial and temporal information together is more efficient comparing to other existing strategies.

Machine learning techniques have excelled in the automatic semantic analysis of images, reaching human-level performances on challenging benchmarks. Yet, the semantic analysis of videos remains challenging due to the significantly higher dimensionality of the input data, respectively, the significantly higher need for annotated training examples. By studying the automatic recognition of German sign language videos, we demonstrate that on the relatively scarce training data of 2.800 videos, modern deep learning architectures for video analysis (such as ResNeXt) along with transfer learning on large gesture recognition tasks, can achieve about 75% character accuracy. Considering that this leaves us with a probability of under 25% that a 5 letter word is spelled correctly, spell-correction systems are crucial for producing readable outputs. The contribution of this paper is to propose a convolutional neural network for spell-correction that expects the softmax outputs of the character recognition network (instead of a misspelled word) as an input. We demonstrate that purely learning on softmax inputs in combination with scarce training data yields overfitting as the network learns the inputs by heart. In contrast, training the network on several variants of the logits of the classification output i.e. scaling by a constant factor, adding of random noise, mixing of softmax and hardmax inputs or purely training on hardmax inputs, leads to better generalization while benefitting from the significant information hidden in these outputs (that have 98% top-5 accuracy), yielding a readable text despite the comparably low character accuracy., First two authors contributed equally. Accepted at ICPR 2020

Domain adaptation (DA) is a representation learning methodology that transfers knowledge from a label-sufficient source domain to a label-scarce target domain. While most of early methods are focused on unsupervised DA (UDA), several studies on semi-supervised DA (SSDA) are recently suggested. In SSDA, a small number of labeled target images are given for training, and the effectiveness of those data is demonstrated by the previous studies. However, the previous SSDA approaches solely adopt those data for embedding ordinary supervised losses, overlooking the potential usefulness of the few yet informative clues. Based on this observation, in this paper, we propose a novel method that further exploits the labeled target images for SSDA. Specifically, we utilize labeled target images to selectively generate pseudo labels for unlabeled target images. In addition, based on the observation that pseudo labels are inevitably noisy, we apply a label noise-robust learning scheme, which progressively updates the network and the set of pseudo labels by turns. Extensive experimental results show that our proposed method outperforms other previous state-of-the-art SSDA methods., Accepted at ICPR 2020

This paper investigates the dependence of existing state-of-the-art person re-identification models on the presence and visibility of human faces. We apply a face detection and blurring algorithm to create anonymized versions of several popular person re-identification datasets including Market1501, DukeMTMC-reID, CUHK03, Viper, and Airport. Using a cross-section of existing state-of-the-art models that range in accuracy and computational efficiency, we evaluate the effect of this anonymization on re-identification performance using standard metrics. Perhaps surprisingly, the effect on mAP is very small, and accuracy is recovered by simply training on the anonymized versions of the data rather than the original data. These findings are consistent across multiple models and datasets. These results indicate that datasets can be safely anonymized by blurring faces without significantly impacting the performance of person reidentification systems, and may allow for the release of new richer re-identification datasets where previously there were privacy or data protection concerns., 25th International Conference on Pattern Recognition (ICPR2020), Milan, Italy, 10-15 January 2021

This paper focuses on the problem of visual saliency prediction, predicting regions of an image that tend to attract human visual attention, under a constrained computational budget. We modify and test various recent efficient convolutional neural network architectures like EfficientNet and MobileNetV2 and compare them with existing state-of-the-art saliency models such as SalGAN and DeepGaze II both in terms of standard accuracy metrics like AUC and NSS, and in terms of the computational complexity and model size. We find that MobileNetV2 makes an excellent backbone for a visual saliency model and can be effective even without a complex decoder. We also show that knowledge transfer from a more computationally expensive model like DeepGaze II can be achieved via pseudo-labelling an unlabelled dataset, and that this approach gives result on-par with many state-of-the-art algorithms with a fraction of the computational cost and model size. Source code is available at https://github.com/feiyanhu/FastSal.

Visual Question Answering (VQA) is challenging due to the complex cross-modal relations. It has received extensive attention from the research community. From the human perspective, to answer a visual question, one needs to read the question and then refer to the image to generate an answer. This answer will then be checked against the question and image again for the final confirmation. In this paper, we mimic this process and propose a fully attention based VQA architecture. Moreover, an answer-checking module is proposed to perform a unified attention on the jointly answer, question and image representation to update the answer. This mimics the human answer checking process to consider the answer in the context. With answer-checking modules and transferred BERT layers, our model achieves the state-of-the-art accuracy 71.57% using fewer parameters on VQA-v2.0 test-standard split.

International audience; In this paper we benchmark different types of cameras and evaluate their performance in terms of reliable localization reliability and precision in Visual Simultaneous Localization and Mapping (vSLAM). Such benchmarking is merely found for visual odometry, but never for vSLAM. Existing studies usually compare several algorithms for a given camera. The evaluation methodology we propose is applied to the recent OpenVSLAM framework. The latter is versatile enough to natively deal with perspective, fisheye, 360 cameras in a monocular or stereoscopic setup, an in RGB or RGB-D modalities. Results in various sequences containing light variation and scenery modifications in the scene assess quantitatively the maximum localization rate for 360 vision. In the contrary, RGB-D vision shows the lowest localization rate, but highest precision when localization is possible. Stereo-fisheye trades-off with localization rates and precision between 360 vision and RGB-D vision. The dataset with ground truth will be made available in open access to allow evaluating other/future vSLAM algorithms with respect to these camera types.

Computer vision with state-of-the-art deep learning models has achieved huge success in the field of Optical Character Recognition (OCR) including text detection and recognition tasks recently. However, Key Information Extraction (KIE) from documents as the downstream task of OCR, having a large number of use scenarios in real-world, remains a challenge because documents not only have textual features extracting from OCR systems but also have semantic visual features that are not fully exploited and play a critical role in KIE. Too little work has been devoted to efficiently make full use of both textual and visual features of the documents. In this paper, we introduce PICK, a framework that is effective and robust in handling complex documents layout for KIE by combining graph learning with graph convolution operation, yielding a richer semantic representation containing the textual and visual features and global layout without ambiguity. Extensive experiments on real-world datasets have been conducted to show that our method outperforms baselines methods by significant margins. Our code is available at https://github.com/wenwenyu/PICK-pytorch., Comment: Accepted by ICPR2020. Code at https://github.com/wenwenyu/PICK-pytorch

With the development of cheap image sensors, the amount of available image data have increased enormously, and the possibility of using crowdsourced collection methods has emerged. This calls for development of ways to handle all these data. In this paper, we present new tools that will enable efficient, flexible and robust map merging. Assuming that separate optimisations have been performed for the individual maps, we show how only relevant data can be stored in a low memory footprint representation. We use these representations to perform map merging so that the algorithm is invariant to the merging order and independent of the choice of coordinate system. The result is a robust algorithm that can be applied to several maps simultaneously. The result of a merge can also be represented with the same type of low-memory footprint format, which enables further merging and updating of the map in a hierarchical way. Furthermore, the method can perform loop closing and also detect changes in the scene between the capture of the different image sequences. Using both simulated and real data - from both a hand held mobile phone and from a drone - we verify the performance of the proposed method., Comment: Accepted at ICPR2020, 9 pages, 8 figures

International audience; With the advent of data living on vertices of graphs, there is much interest in processing the so-called graph signals for partitioning tasks. As active contours have had much impact in the image processing community, their formulation on graphs is of importance to the field of graph signal processing. This paper proposes an adaptation on graphs of a model that combines the Geodesic Active Contour and the Active Contour Without Edges models. In addition, specific terms depending on graphs are introduced in the formulation. This adaptation is solved using a level set formulation with a gradient descent that can be expressed as a morphological front evolution process. Experimental results on different kinds of graphs signals show the benefit of the approach.

The large spread of online shopping has led computer vision researchers to develop different solutions for the fashion domain to potentially increase the online user experience and improve the efficiency of preparing fashion catalogs. Among them, image-based virtual try-on has recently attracted a lot of attention resulting in several architectures that can generate a new image of a person wearing an input try-on garment in a plausible and realistic way. In this paper, we present VITON-G T, a new model for virtual try-on that generates high-quality and photo-realistic images thanks to multiple geometric transformations. In particular, our model is composed of a two-stage geometric transformation module that performs two different projections on the input garment, and a transformation-guided try-on module that synthesizes the new image. We experimentally validate the proposed solution on the most common dataset for this task, containing mainly t-shirts, and we demonstrate its effectiveness compared to different baselines and previous methods. Additionally, we assess the generalization capabilities of our model on a new set of fashion items composed of upper-body clothes from different categories. To the best of our knowledge, we are the first to test virtual try-on architectures in this challenging experimental setting.

We propose a new task named Audio-driven Performance Video Generation (APVG), which aims to synthesize the video of a person playing a certain instrument guided by a given music audio clip. It is a challenging task to generate the high-dimensional temporal consistent videos from low-dimensional audio modality. In this paper, we propose a multi-staged framework to generate realistic and synchronized performance video from given music. Firstly, we provide both global appearance and local spatial information by generating the coarse videos and keypoints of body and hands from a given music respectively. Then, we propose to transform the generated keypoints to heatmap via a differentiable space transformer, since the heatmap provides more spatial information but is harder to generate directly from audio. Finally, we propose a Structured Temporal UNet (STU) to extract both intra-frame structured information and interframe temporal consistency. They are obtained via graph-based structure module, and CNN-GRU based high-level temporal module respectively for final video generation. Comprehensive experiments validate the effectiveness of our proposed framework.

Capsules are grouping of neurons that allow to represent sophisticated information of a visual entity such as pose and features. In the view of this property, Capsule Networks outperform CNNs in challenging tasks like object recognition in unseen viewpoints, and this is achieved by learning the transformations between the object and its parts with the help of high dimensional representation of pose information. In this paper, we present Quaternion Capsules (QCN) where pose information of capsules and their transformations are represented by quaternions. Quaternions are immune to the gimbal lock, have straightforward regularization of the rotation representation for capsules, and require less number of parameters than matrices. The experimental results show that QCNs generalize better to novel viewpoints with fewer parameters, and also achieve on-par or better performances with the state-of-the-art Capsule architectures on well-known benchmarking datasets. Our code is available11https://github.com/Boazrciasn/Quaternion-Capsule-Networks.git.

In music analysis, one of the most fundamental tasks is note onset detection - detecting the beginning of new note events. As the target function of onset detection is related to other tasks, such as beat tracking or tempo estimation, onset detection is the basis for such related tasks. Furthermore, it can help to improve Automatic Music Transcription (AMT). Typically, different approaches for onset detection follow a similar outline: An audio signal is transformed into an Onset Detection Function (ODF), which should have rather low values (i.e. close to zero) for most of the time but with pronounced peaks at onset times, which can then be extracted by applying peak picking algorithms on the ODF. In the recent years, several kinds of neural networks were used successfully to compute the ODF from feature vectors. Currently, Convolutional Neural Networks (CNNs) define the state of the art. In this paper, we build up on an alternative approach to obtain a ODF by Echo State Networks (ESNs), which have achieved comparable results to CNNs in several tasks, such as speech and image recognition. In contrast to the typical iterative training procedures of deep learning architectures, such as CNNs or networks consisting of Long-Short-Term Memory Cells (LSTMs), in ESNs only a very small part of the weights is easily trained in one shot using linear regression. By comparing the performance of several feature extraction methods, pre-processing steps and introducing a new way to stack ESNs, we expand our previous approach to achieve results that fall between a bidirectional LSTM network and a CNN with relative improvements of 1.8 % and -1.4 %, respectively. For the evaluation, we used exactly the same 8-fold cross validation setup as for the reference results.

Although deep neural networks (NNs) have achieved state-of-the-art accuracy in many visual recognition tasks, the growing computational complexity and energy consumption of networks remains an issue, especially for applications on platforms with limited resources and requiring real-time processing. Filter pruning techniques have recently shown promising results for the compression and acceleration of convolutional NNs (CNNs). However, these techniques involve numerous steps and complex optimisations because some only prune after training CNNs, while others prune from scratch during training, by integrating sparsity constraints or by modifying the loss function. In this paper, we introduce a new Progressive Gradient Pruning (PGP) technique for iterative filter pruning during training. In contrast to previous progressive pruning techniques, it relies on a novel filter selection criterion that measures the change in filter weights, uses a new hard and soft pruning strategy, and effectively adapts momentum tensors during the backward propagation pass. Experimental results obtained after training various CNNs on benchmark datasets for image classification, object detection and domain adaptation indicate that our PGP technique can achieve a better trade-off between classification accuracy and network (time and memory) complexity than PSFP and other state-of-the-art filter pruning techniques. Code is available on GitHub link: https://github.com/Anon6627/Pruning-PGP.

Event-based cameras are vision devices that transmit only brightness changes with low latency and ultra-low power consumption. Such characteristics make event-based cameras attractive in the field of localization and object tracking in resource-constrained systems. Since the number of generated events in such cameras is huge, the selection and filtering of the incoming events are beneficial from both increasing the accuracy of the features and reducing the computational load. In this paper, we present an algorithm to detect asynchronous corners from a stream of events in real-time on embedded systems. The algorithm is called the Three Layer Filtering-Harris or TLF-Harris algorithm. The algorithm is based on an events' filtering strategy whose purpose is 1) to increase the accuracy by deliberately eliminating some incoming events, i.e., noise, and 2) to improve the real-time performance of the system, i.e., preserving a constant throughput in terms of input events per second, by discarding unnecessary events with a limited accuracy loss. An approximation of the Harris algorithm, in turn, is used to exploit its high-quality detection capability with a low-complexity implementation to enable seamless real-time performance on embedded computing platforms. The proposed algorithm is capable of selecting the best corner candidate among neighbors and achieves an average execution time savings of 59 % compared with the conventional Harris score. Moreover, our approach outperforms the competing methods, such as eFAST, eHarris, and FA-Harris, in terms of real-time performance, and surpasses Arc* in terms of accuracy., Accepted to 25th international conference on pattern recognition (ICPR 2020)

Different techniques have emerged in the deep learning scenario, such as Convolutional Neural Networks, Deep Belief Networks, and Long Short-Term Memory Networks, to cite a few. In lockstep, regularization methods, which aim to prevent overfitting by penalizing the weight connections, or turning off some units, have been widely studied either. In this paper, we present a novel approach called MaxDropout, a regularizer for deep neural network models that works in a supervised fashion by removing (shutting off) the prominent neurons (i.e., most active) in each hidden layer. The model forces fewer activated units to learn more representative information, thus providing sparsity. Regarding the experiments, we show that it is possible to improve existing neural networks and provide better results in neural networks when Dropout is replaced by MaxDropout. The proposed method was evaluated in image classification, achieving comparable results to existing regularizers, such as Cutout and RandomErasing, also improving the accuracy of neural networks that uses Dropout by replacing the existing layer by MaxDropout.

The availability of large-scale annotated datasets has enabled Fully-Convolutional Neural Networks to reach outstanding performance on road extraction in aerial images. However, high-quality pixel-level annotation is expensive to produce and even manually labeled data often contains topological errors. Trading off quality for quantity, many datasets rely on already available yet noisy labels, for example from OpenStreetMap. In this paper, we explore the training of custom U-Nets built with ResNet and DenseNet backbones using noise-aware losses that are robust towards label omission and registration noise. We perform an extensive evaluation of standard and noise-aware losses, including a novel Bootstrapped DICE-Coefficient loss, on two challenging road segmentation benchmarks. Our losses yield a consistent improvement in overall extraction quality and exhibit a strong capacity to cope with severe label noise. Our method generalizes well to two other fine-grained topology delineation tasks: surface crack detection for quality inspection and cell membrane extraction in electron microscopy imagery.

To account for appearance variations, tracking models need to be updated during the course of inference. However, updating the tracker model with adverse bounding box predictions adds an unavoidable bias term to the learning. This bias term, which we refer to as model decay, offsets the learning and causes tracking drift. While its adverse affect might not be visible in short-term tracking, accumulation of this bias over a long-term can eventually lead to a permanent loss of the target. In this paper, we look at the problem of model bias from a mathematical perspective. Further, we briefly examine the effect of various sources of tracking error on model decay, using a correlation filter (ECO) and a Siamese (SINT) tracker. Based on observations and insights, we propose simple additions that help to reduce model decay in long-term tracking. The proposed tracker is evaluated on four long-term and one short-term tracking benchmarks, demonstrating superior accuracy and robustness, even on 30 minute long videos.

Camera pose estimation in large-scale environments is still an open question and, despite recent promising results, it may still fail in some situations. The research so far has focused on improving subcomponents of estimation pipelines, to achieve more accurate poses. However, there is no guarantee for the result to be correct, even though the correctness of pose estimation is critically important in several visual localization applications,such as in autonomous navigation. In this paper we bring to attention a novel research question, pose confidence estimation,where we aim at quantifying how reliable the visually estimated pose is. We develop a novel confidence measure to fulfil this task and show that it can be flexibly applied to different datasets,indoor or outdoor, and for various visual localization pipelines.We also show that the proposed techniques can be used to accomplish a secondary goal: improving the accuracy of existing pose estimation pipelines. Finally, the proposed approach is computationally light-weight and adds only a negligible increase to the computational effort of pose estimation., Comment: To appear in ICPR 2020

In this paper, we study the task of facial expression recognition under strong occlusion. We are particularly interested in cases where 50% of the face is occluded, e.g. when the subject wears a Virtual Reality (VR) headset. While previous studies show that pre-training convolutional neural networks (CNNs) on fully-visible (non-occluded) faces improves the accuracy, we propose to employ knowledge distillation to achieve further improvements. First of all, we employ the classic teacher-student training strategy, in which the teacher is a CNN trained on fully-visible faces and the student is a CNN trained on occluded faces. Second of all, we propose a new approach for knowledge distillation based on triplet loss. During training, the goal is to reduce the distance between an anchor embedding, produced by a student CNN that takes occluded faces as input, and a positive embedding (from the same class as the anchor), produced by a teacher CNN trained on fully-visible faces, so that it becomes smaller than the distance between the anchor and a negative embedding (from a different class than the anchor), produced by the student CNN. Third of all, we propose to combine the distilled embeddings obtained through the classic teacher-student strategy and our novel teacher-student strategy based on triplet loss into a single embedding vector. We conduct experiments on two benchmarks, FER+ and AffectNet, with two CNN architectures, VGG-f and VGG-face, showing that knowledge distillation can bring significant improvements over the state-of-the-art methods designed for occluded faces in the VR setting., Accepted at ICPR 2020

With expansion of the video advertising market, research to predict the effects of video advertising is getting more attention. Although effect prediction of image advertising has been explored a lot, prediction for video advertising is still challenging with seldom research. In this research, we propose a method for predicting the click through rate (CTR) of video advertisements and analyzing the factors that determine the CTR. In this paper, we demonstrate an optimized framework for accurately predicting the effects by taking advantage of the multimodal nature of online video advertisements including video, text, and metadata features. In particular, the two types of metadata, i.e., categorical and continuous, are properly separated and normalized. To avoid overfitting, which is crucial in our task because the training data are not very rich, additional regularization layers are inserted. Experimental results show that our approach can achieve a correlation coefficient as high as 0.695, which is a significant improvement from the baseline (0.487)., Comment: Accepted at International Conference on Pattern Recognition 2020 (ICPR)

Medical image segmentation has played an important role in medical analysis and widely developed for many clinical applications. Deep learning-based approaches have achieved high performance in semantic segmentation but they are limited to pixel-wise setting and imbalanced classes data problem. In this paper, we tackle those limitations by developing a new deep learning-based model which takes into account both higher feature level i.e. region inside contour, intermediate feature level i.e. offset curves around the contour and lower feature level i.e. contour. Our proposed Offset Curves (OsC) loss consists of three main fitting terms. The first fitting term focuses on pixel-wise level segmentation whereas the second fitting term acts as attention model which pays attention to the area around the boundaries (offset curves). The third terms plays a role as regularization term which takes the length of boundaries into account. We evaluate our proposed OsC loss on both 2D network and 3D network. Two common medical datasets, i.e. retina DRIVE and brain tumor BRATS 2018 datasets are used to benchmark our proposed loss performance. The experiments have shown that our proposed OsC loss function outperforms other mainstream loss functions such as Cross-Entropy, Dice, Focal on the most common segmentation networks Unet, FCN., ICPR 2020

The demands of automated shipping address recognition and verification have increased to handle a large number of packages and to save costs associated with misdelivery. A previous study proposed a deep learning system where the shipping address is recognized and verified based on a camera image capturing the shipping address and barcode area. Because the system performance depends on the input image quality, inspection of input image quality is necessary for image preprocessing. In this paper, we propose an input image quality verification method combining global and local features. Object detection and scale-invariant feature transform in different feature spaces are developed to extract global and local features from several independent convolutional neural networks. The conditions of shipping label images are classified by fully connected fusion layers with concatenated global and local features. The experimental results regarding real captured and generated images show that the proposed method achieves better performance than other methods. These results are expected to improve the shipping address recognition and verification system by applying different image preprocessing steps based on the classified conditions., Accepted at ICPR 2020

International audience; The accurate knowledge of vascular network geometry is crucial for many clinical applications such as cardiovascular disease diagnosis and surgery planning. Vessel enhancement algorithms are often a key step to improve the robustness of vessel segmentation. A wide variety of enhancement filters exists in the literature, but they are often difficult to compare as the applications and datasets differ from a paper to another and the code is rarely available. In this article, we compare seven vessel enhancement filters covering the last twenty years literature in a unique common framework. We focus our study on the liver vascular network which is under-represented in the literature. The evaluation is made from three points of view: in the whole liver, in the vessel neighborhood and near the bifurcations. The study is performed on two publicly available datasets: the Ircad dataset (CT images) and the VascuSynth dataset adapted for MRI simulation. We discuss the strengths and weaknesses of each method in the hepatic context. In addition, the benchmark framework including a C++ implementation of each compared method is provided. An online demonstration ensures the reproducibility of the results without requiring any additional software.

In this paper, a new 2D transform named Discrete Mirror Transform (DMT) is presented. The DMT is computed by decomposing a signal into its even and odd parts around an optimal location in a given direction so that the signal energy is maximally split between the two components. After minimizing the information required to regenerate the original signal by removing redundant structures, the process is iterated leading the signal energy to distribute into a continuously smaller set of coefficients. The DMT can be displayed as a binary tree, where each node represents the single (even or odd) signal derived from the decomposition in the previous level. An optimized version of the DMT (ODMT) is also introduced, by exploiting the possibility to choose different directions at which performing the decomposition. Experimental simulations have been carried out in order to test the sparsity properties of the DMT and ODMT when applied on images: referring to both transforms, the results show a superior performance with respect to the popular Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) in terms of non-linear approximation.

In this paper, we propose an automatic approach for localizing the inner eye canthus in thermal face images. We first coarsely detect 5 facial keypoints corresponding to the center of the eyes, the nosetip and the ears. Then we compute a sparse 2D-3D points correspondence using a 3D Morphable Face Model (3DMM). This correspondence is used to project the entire 3D face onto the image, and subsequently locate the inner eye canthus. Detecting this location allows to obtain the most precise body temperature measurement for a person using a thermal camera. We evaluated the approach on a thermal face dataset provided with manually annotated landmarks. However, such manual annotations are normally conceived to identify facial parts such as eyes, nose and mouth, and are not specifically tailored for localizing the eye canthus region. As additional contribution, we enrich the original dataset by using the annotated landmarks to deform and project the 3DMM onto the images. Then, by manually selecting a small region corresponding to the eye canthus, we enrich the dataset with additional annotations. By using the manual landmarks, we ensure the correctness of the 3DMM projection, which can be used as ground-truth for future evaluations. Moreover, we supply the dataset with the 3D head poses and per-point visibility masks for detecting self-occlusions. The data is publicly available at https://www.micc.unifi.it/resources/datasets/thermal-face/.