Although deeper and larger neural networks have achieved better performance, due to overwhelming burden on computation, they cannot meet the demands of deployment on resource-limited devices. An effective strategy to address this problem is to make use of dynamic inference mechanism, which changes the inference path for different samples at runtime. Existing methods only reduce the depth by skipping an entire specific layer, which may lose important information in this layer. In this paper, we propose a novel method called Dynamic Multi-path Neural Network (DMNN), which provides more topology choices in terms of both width and depth on the fly. For better modelling the inference path selection, we further introduce previous state and object category information to guide the training process. Compared to previous dynamic inference techniques, the proposed method is more flexible and easier to incorporate into most modern network architectures. Experimental results on ImageNet and CIFAR-100 demonstrate the superiority of our method on both efficiency and classification accuracy.

Interpretable Machine Learning (IML) methods are used to gain insight into the relevance of a feature of interest for the performance of a model. Commonly used IML methods differ in whether they consider features of interest in isolation, e.g., Permutation Feature Importance (PFI), or in relation to all remaining feature variables, e.g., Conditional Feature Importance (CFI). As such, the perturbation mechanisms inherent to PFI and CFI represent extreme reference points. We introduce Relative Feature Importance (RFI), a generalization of PFI and CFI that allows for a more nuanced feature importance computation beyond the PFI versus CFI dichotomy. With RFI, the importance of a feature relative to any other subset of features can be assessed, including variables that were not available at training time. We derive general interpretation rules for RFI based on a detailed theoretical analysis of the implications of relative feature relevance, and demonstrate the method's usefulness on simulated examples.

In this work we propose a deep learning pipeline to predict the visual future appearance of an urban scene. Despite recent advances, generating the entire scene in an end-to-end fashion is still far from being achieved. Instead, here we follow a two stages approach, where interpretable information is included in the loop and each actor is modelled independently. We leverage a per-object novel view synthesis paradigm; i.e. generating a synthetic representation of an object undergoing a geometrical roto-translation in the 3D space. Our model can be easily conditioned with constraints (e.g. input trajectories) provided by state-of-the-art tracking methods or by the user itself. This allows us to generate a set of diverse realistic futures starting from the same input in a multi-modal fashion. We visually and quantitatively show the superiority of this approach over traditional end-to-end scene-generation methods on CityFlow, a challenging real world dataset.

Deep neural networks (DNNs) with many layers of feature representations yield state-of-the-art results on several difficult learning tasks. However, optimizing very deep DNNs without shortcut connections known as PlainNets, is a notoriously hard problem. Considering the growing interest in this area, this paper investigates holistically two scenarios that plague the training of very deep PlainNets: (1) the relatively popular challenge of 'vanishing and exploding units' activations', and (2) the less investigated 'singularity' problem, which is studied in details in this paper. In contrast to earlier works that study only the saturation and explosion of units' activations in isolation, this paper harmonizes the inconspicuous coexistence of the aforementioned problems for very deep PlainNets. Particularly, we argue that the aforementioned problems would have to be tackled simultaneously for the successful training of very deep PlainNets. Finally, different techniques that can be employed for tackling the optimization problem are discussed, and a specific combination of simple techniques that allows the successful training of PlainNets having up to 100 layers is demonstrated.

In this paper, we present a novel framework that uses tensor factorization to generate richer feature spaces for pixel classification in hyperspectral images. In particular, we assess the performance of different tensor rank decomposition methods as compared to the traditional kernel-based approaches for the hyperspectral image classification problem. We propose ORION, which takes as input a hyperspectral image tensor and a rank and outputs an enhanced feature space from the factor matrices of the decomposed tensor. Our method is a feature explosion technique that inherently maps low dimensional input space in R^K to high dimensional space in R^R, where R >> K, say in the order of 1000x, like a kernel. We show how the proposed method exploits the multi-linear structure of hyperspectral three-dimensional tensor. We demonstrate the effectiveness of our method with experiments on three publicly available hyperspectral datasets with labeled pixels and compare their classification performance against traditional linear and non-linear supervised learning methods such as SVM with Linear, Polynomial, RBF kernels, and the Multi-Layer Perceptron model. Finally, we explore the relationship between the rank of the tensor decomposition and the classification accuracy using several hyperspectral datasets with ground truth.

In recent years, deep neural networks (DNN) have been applied to different domains and achieved dramatic performance improvements over state-of-the-art classical methods. These performances of DNNs were however often obtained with networks containing millions of parameters and which training required heavy computational power. In order to cope with this computational issue a huge literature deals with proximal regularization methods which are time consuming.\\ In this paper, we propose instead a constrained approach. We provide the general framework for our new splitting projection gradient method. Our splitting algorithm iterates a gradient step and a projection on convex sets. We study algorithms for different constraints: the classical $\ell_1$ unstructured constraint and structured constraints such as the nuclear norm, the $\ell_{2,1} $ constraint (Group LASSO). We propose a new $\ell_{1,1} $ structured constraint for which we provide a new projection algorithm We demonstrate the effectiveness of our method on three popular datasets (MNIST, Fashion MNIST and CIFAR). Experiments on these datasets show that our splitting projection method with our new $\ell_{1,1} $ structured constraint provides the best reduction of memory and computational power. Experiments show that fully connected linear DNN are more efficient for green AI.

In this paper, we proposed two modified neural networks based on dual path multi-scale fusion networks (SFANet) and SegNet for accurate and efficient crowd counting. Inspired by SFANet, the first model, which is named M-SFANet, is attached with atrous spatial pyramid pooling (ASPP) and context-aware module (CAN). The encoder of M-SFANet is enhanced with ASPP containing parallel atrous convolutional layers with different sampling rates and hence able to extract multi-scale features of the target object and incorporate larger context. To further deal with scale variation throughout an input image, we leverage the CAN module which adaptively encodes the scales of the contextual information. The combination yields an effective model for counting in both dense and sparse crowd scenes. Based on the SFANet decoder structure, M-SFANet's decoder has dual paths, for density map and attention map generation. The second model is called M-SegNet, which is produced by replacing the bilinear upsampling in SFANet with max unpooling that is used in SegNet. This change provides a faster model while providing competitive counting performance. Designed for high-speed surveillance applications, M-SegNet has no additional multi-scale-aware module in order to not increase the complexity. Both models are encoder-decoder based architectures and are end-to-end trainable. We conduct extensive experiments on five crowd counting datasets and one vehicle counting dataset to show that these modifications yield algorithms that could improve state-of-the-art crowd counting methods.

In the field of multimodal segmentation, the correlation between different modalities can be considered for improving the segmentation results. In this paper, we propose a multi-modality segmentation network with a correlation constraint. Our network includes N model-independent encoding paths with N image sources, a correlation constrain block, a feature fusion block, and a decoding path. The model-independent encoding path can capture modality-specific features from the N modalities. Since there exists a strong correlation between different modalities, we first propose a linear correlation block to learn the correlation between modalities, then a loss function is used to guide the network to learn the correlated features based on the correlation representation block. This block forces the network to learn the latent correlated features which are more relevant for segmentation. Considering that not all the features extracted from the encoders are useful for segmentation, we propose to use dual attention based fusion block to recalibrate the features along the modality and spatial paths, which can suppress less informative features and emphasize the useful ones. The fused feature representation is finally projected by the decoder to obtain the segmentation result. Our experiment results tested on BraTS-2018 dataset for brain tumor segmentation demonstrate the effectiveness of our proposed method.

Recent machine learning methods use increasingly large deep neural networks to achieve state-of-the-art results in various tasks. Network quantization can effectively reduce computation and memory costs without modifying network structures, facilitating the deployment of deep neural networks (DNNs) on cloud and edge devices. However, most of the existing methods usually need time-consuming training or fine-tuning and access to the original training dataset that may be unavailable due to privacy or security concerns. In this paper, we present a novel method to achieve low-precision quantization of deep neural networks with limited data. Firstly, to reduce the complexity of per-channel quantization and the degeneration of per-layer quantization, we introduce group-wise quantization which separates the output channels into groups that each group is quantized separately. Secondly, to better distill knowledge from the pre-trained FP32 model with limited data, we introduce a two-stage knowledge distillation method that divides the optimization process into independent optimization stage and joint optimization stage to address the limitation of layer-wise supervision and global supervision. Extensive experiments on ImageNet2012 (ResNet18/50, ShuffleNetV2, and MobileNetV2) demonstrate that the proposed approach can significantly improve the quantization model's accuracy when only a few training samples are available. We further show that the method also extends to other computer vision architectures and tasks such as object detection and semantic segmentation.

Person re-identification (Re-ID) has attracted a lot of research attention in recent years. However, supervised methods demand an enormous amount of manually annotated data. In this paper, we propose a Self-Paced bottom-up Clustering Network with Side Information (SPCNet-SI) for unsupervised person Re-ID, where the side information comes from the serial number of the camera associated with each image. Specifically, our proposed SPCNet-SI exploits the camera side information to guide the feature learning and uses soft label in bottom-up clustering process, in which the camera association information is used in the repelled loss and the soft label based cluster information is used to select the candidate cluster pairs to merge. Moreover, a self-paced dynamic mechanism is developed to regularize the merging process such that the clustering is implemented in an easy-to-hard way with a slow-to-fast merging process. Experiments on two benchmark datasets Market-1501 and DukeMTMC-ReID demonstrate promising performance.

This work addresses the pose estimation task on low-resolution images captured using thermal sensors which can operate in a no-light environment. Low-resolution thermal sensors have been widely adopted in various applications for cost control and privacy protection purposes. In this paper, targeting the challenging scenario of ultra-low resolution thermal imaging (3232 pixels), we aim to estimate human poses for the purpose of monitoring health conditions and indoor events. To overcome the challenges in ultra-low resolution thermal imaging such as blurred boundaries and data scarcity, we propose a new Image-to-Image (I2I) translation architecture which can translate the original blurred thermal image into a visible light image with sharper boundaries. Then the generated visible light image can be fed into the off-the-shelf pose estimator which was well-trained in the visible domain. Experimental results suggest that the proposed framework outperforms other state-of-the-art methods in the I2I based pose estimation task for our thermal image dataset. Furthermore, we also demonstrated the merits of the proposed method on the publicly available FLIR dataset by measuring the quality of translated images.

Attributed graphs are structures that are useful to represent objects through the information of their local parts and their relations. Each characteristic in the local parts is represented by different attributes on the nodes. In this context, the comparison between structured objects is performed through a distance between attributed graphs. If we want to correctly tune the distance and the node correspondence between graphs, we have to add some weights on the node attributes to gauge the importance of each local characteristic. In this paper, we present a method to learn the weights on each node attribute. It is based on building an embedded space and imposing the weights we want to learn to be the constants of the hyperplane deduced by a linear regression applied on a cloud of points. These points represent the node-to-node mappings.

This paper presents a novel deep learning framework for the electrocardiogram (ECG) heartbeat classification. Although there have been some studies with excellent overall accuracy, these studies have not been very accurate in the diagnosis of arrhythmia classes especially such as supraventricular ectopic beat (SVEB) and ventricular ectopic beat (VEB). In our work, we propose a Multi-Scale and Attention based ResNet for heartbeat classification in intra-patient and inter-patient paradigms respectively. Firstly, we extract shallow features from a convolutional layer. Secondly, the shallow features are sent into three branches with different convolution kernels in order to combine receptive fields of different sizes. Finally, fully connected layers are used to classify the heartbeat. Besides, we design a new attention mechanism based on the characteristics of heartbeat data. At last, extensive experiments on benchmark dataset demonstrate the effectiveness of our proposed model.

Spherical cameras and the latest image processing techniques open up new horizons. In particular, methods based on Convolutional Neural Networks (CNNs) now give excellent results for optical flow estimation on perspective images. However, these approaches are highly dependent on their architectures and training datasets. This paper proposes to benefit from years of improvement in perspective images optical flow estimation and to apply it to omnidirectional ones without training on new datasets. Our network, OmniFlowNet, is built on a CNN specialized in perspective images. Its convolution operation is adapted to be consistent with the equirectangular projection. Tested on spherical datasets created with Blender and several equirectangular videos realized from real indoor and outdoor scenes, OmniFlowNet shows better performance than its original network.

Mixture Model (MM) is a probabilistic framework allows us to define dataset containing $K$ different modes. When each of the modes is associated with a Gaussian distribution, we refer to it as Gaussian MM or GMM. Given a data point $x$, a GMM may assume the existence of a random index $k \in \{1, \dots , K \}$ identifying which Gaussian the particular data is associated with. In a traditional GMM paradigm, it is straightforward to compute in closed-form, the conditional likelihood $p(x |k, \theta)$ as well as the responsibility probability $p(k|x, \theta)$ describing the distribution weights for each data. Computing the responsibility allows us to retrieve many important statistics of the overall dataset, including the weights of each of the modes/clusters. Modern large data-sets are often containing multiple unlabelled modes, such as paintings dataset may contain several styles; fashion images containing several unlabelled categories. In its raw representation, the Euclidean distances between the data (e.g., images) do not allow them to form mixtures naturally, nor it's feasible to compute responsibility distribution analytically, making GMM unable to apply. In this paper, we utilize the Generative Adversarial Network (GAN) framework to achieve a plausible alternative method to compute these probabilities. The key insight is that we compute them at the data's latent space $z$ instead of $x$. However, this process of $z \rightarrow x$ is irreversible under GAN which renders the computation of responsibility $p(k|x, \theta)$ infeasible. Our paper proposed a novel method to solve it by using a so-called Posterior Consistency Module (PCM). PCM acts like a GAN, except its Generator $C_{\text{PCM}}$ does not output the data, but instead it outputs a distribution to approximate $p(k|x, \theta)$. The entire network is trained in an ``end-to-end'' fashion. Trough these techniques, it allows us to model the dataset of very complex structure using GMM and subsequently to discover interesting properties of an unsupervised dataset, including its segments, as well as generating new ``out-distribution" data by smooth linear interpolation across any combinations of the modes in a completely unsupervised manner.

Creating a good looking chart for better visualization is time consuming. There are plenty of well-designed charts on the Web, which are ideal references for imitation of chart style. However, stored as bitmap images, reference charts have hinder machine interpretation of style settings and thus difficult to be directly applied. In this paper, we extract visual properties from reference chart images as style templates to restyle charts. We first construct a large-scale dataset of 187,059 chart images from real world data, labeled with predefined visual property values. Then we introduce an end-to-end learning network to extract the properties based on two image-encoding approaches. Furthermore, in order to capture spatial relationships of chart objects, which are crucial in solving the task, we propose a novel positional encoding method to integrate clues of relative positions between objects. Experimental results show that our model significantly outperforms baseline models. By adding positional features, our model achieves better performance. Finally, we present the application for chart restyling based on our model.

One of the significant challenges of deep neural networks is that the complex nature of the network prevents human comprehension of the outcome of the network. Consequently, the applicability of complex machine learning models is limited in the safety-critical domains, which incurs risk to life and property. To fully exploit the capabilities of complex neural networks, we propose a non-intrusive interpretability technique that uses the input and output of the model to generate a saliency map. The method works by empirically optimizing a randomly initialized input mask by localizing and weighing individual pixels according to their sensitivity towards the target class. Our experiments show that the proposed model interpretability approach performs better than the existing saliency map-based approaches methods at localizing the relevant input pixels. Furthermore, to obtain a global perspective on the target-specific explanation, we propose a saliency map reconstruction approach to generate acceptable variations of the salient inputs from the space of input data distribution for which the model outcome remains unaltered. Experiments show that our interpretability method can reconstruct the salient part of the input with a classification accuracy of 89%.

Objects from correlated classes usually share highly similar appearances while objects from uncorrelated classes are very different. Most of current image recognition works treat each class independently, which ignores these class correlations and inevitably leads to sub-optimal performance in many cases. Fortunately, object classes inherently form a hierarchy with different levels of abstraction and this hierarchy encodes rich correlations among different classes. In this work, we utilize a soft label vector that encodes the prior knowledge of class correlations as extra regularization to train the image classifiers. Specifically, for each class, instead of simply using a one-hot vector, we assign a high value to its correlated classes and assign small values to those uncorrelated ones, thus generating knowledge embedded soft labels. We conduct experiments on both general and fine-grained image recognition benchmarks and demonstrate its superiority compared with existing methods.

Foreground segmentation technologies play an important role in applications such as free-viewpoint video (FVV) and sports video analysis. In this situation, we propose a new method that achieves accurate foreground silhouette extraction using dynamic object presence probability (DOPP). Our main contributions are as follows. 1) Object presence probability for each pixel is calculated from the object recognition results based on deep learning. After that, background subtraction is implemented by changing the threshold and the update rate of the background model in response to the object presence probability. Parameter tuning of background subtraction is executed by using the object recognition results to improve the silhouette extraction quality. 2) To calculate more accurate silhouette images, parameters of background subtraction are adjusted by monitoring optical flows between consecutive frames. The object presence probability of the current frame is dynamically updated by using the object presence probability of the previous frame with optical flows. In the experiments, we confirmed that the proposed method achieved more accurate silhouette extraction than conventional methods in three sports sequences.

Forensic analysis is used to detect image forgeries e.g. the copy move forgery and the object removal forgery, respectively. Counter forensic techniques (aka anti-forensic methods to fool the forensic analyst by concealing traces of manipulation) have become popular in the game of cat and mouse between the analyst and the attacker. Classical anti-forensic techniques targeting on SIFT keypoints have been established with particular emphasis on keypoint removal in the context of copy move forgery detection. In this paper we propose a forensic approach countering SIFT keypoint removal by changing to a different type of keypoints in forensic analysis, clearly demonstrating benefits over traditional SIFT keypoint oriented techniques.

We present a simple but yet effective method for learning distinctive 3D local deep descriptors (DIPs) that can be used to register point clouds without requiring an initial alignment. Point cloud patches are extracted, canonicalised with respect to their estimated local reference frame and encoded into rotation-invariant compact descriptors by a PointNet-based deep neural network. DIPs can effectively generalise across different sensor modalities because they are learnt end-to-end from locally and randomly sampled points. Moreover, because DIPs encode only local geometric information, they are robust to clutter, occlusions and missing regions. We evaluate and compare DIPs against alternative hand-crafted and deep descriptors on several indoor and outdoor datasets reconstructed using different sensors. Results show that DIPs (i) achieve comparable results to the state-of-the-art on RGB-D indoor scenes (3DMatch dataset), (ii) outperform state-of-the-art by a large margin on laser-scanner outdoor scenes (ETH dataset), and (iii) generalise to indoor scenes reconstructed with the Visual-SLAM system of Android ARCore.