Deep neural networks with batch normalization (BN-DNNs) are invariant to weight rescaling due to their normalization operations. However, using weight decay (WD) benefits these weight-scale-invariant networks, which is often attributed to an increase of the effective learning rate when the weight norms are decreased. In this paper, we demonstrate the insufficiency of the previous explanation and investigate the implicit biases of stochastic gradient descent (SGD) on BN-DNNs to provide a theoretical explanation for the efficacy of weight decay. We identity two implicit biases of SGD on BN-DNNs: 1) the weight norms in SGD training remain constant in the continuous-time domain and keep increasing in the discrete-time domain; 2) SGD optimizes weight vectors in fully-connected networks or convolution kernels in convolution neural networks by updating components lying in the input feature span, while leaving those components orthogonal to the input feature span unchanged. Thus, SGD without WD accumulates weight noise orthogonal to the input feature span, and cannot eliminate such noise. Our empirical studies corroborate the hypothesis that weight decay suppresses weight noise that is left untouched by SGD. Furthermore, we propose to use weight rescaling (WRS) instead of weight decay to achieve the same regularization effect, while avoiding performance degradation of WD on some momentum-based optimizers. Our empirical results on image recognition show that regardless of optimization methods and network architectures, training BN-DNNs using WRS achieves similar or better performance compared with using WD. We also show that training with WRS generalizes better compared to WD, on other computer vision tasks.
Understanding a scene by decoding the visual relationships depicted in an image has been a long studied problem. While the recent advances in deep learning and the usage of deep neural networks have achieved near human accuracy on many tasks, there still exists a pretty big gap between human and machine level performance when it comes to various visual relationship detection tasks. Developing on earlier tasks like object recognition, segmentation and captioning which focused on a relatively coarser image understanding, newer tasks have been introduced recently to deal with a finer level of image understanding. A Scene Graph is one such technique to better represent a scene and the various relationships present in it. With its wide number of applications in various tasks like Visual Question Answering, Semantic Image Retrieval, Image Generation, among many others, it has proved to be a useful tool for deeper and better visual relationship understanding. In this paper, we present a detailed survey on the various techniques for scene graph generation, their efficacy to represent visual relationships and how it has been used to solve various downstream tasks. We also attempt to analyze the various future directions in which the field might advance in the future. Being one of the first papers to give a detailed survey on this topic, we also hope to give a succinct introduction to scene graphs, and guide practitioners while developing approaches for their applications.
Neural ordinary differential equations (NODEs) treat computation of intermediate feature vectors as trajectories of ordinary differential equation parameterized by a neural network. In this paper, we propose a novel model, delay differential neural networks (DDNN), inspired by delay differential equations (DDEs). The proposed model considers the derivative of the hidden feature vector as a function of the current feature vector and past feature vectors (history). The function is modelled as a neural network and consequently, it leads to continuous depth alternatives to many recent ResNet variants. We propose two different DDNN architectures, depending on the way current and past feature vectors are considered. For training DDNNs, we provide a memory-efficient adjoint method for computing gradients and back-propagate through the network. DDNN improves the data efficiency of NODE by further reducing the number of parameters without affecting the generalization performance. Experiments conducted on synthetic and real-world image classification datasets such as Cifar10 and Cifar100 show the effectiveness of the proposed models.
As the class size grows, maintaining a balanced dataset across many classes is challenging because the data are long-tailed in nature; it is even impossible when the sample-of-interest co-exists with each other in one collectable unit, e.g., multiple visual instances in one image. Therefore, long-tailed classification is the key to deep learning at scale. However, existing methods are mainly based on re-weighting/re-sampling heuristics that lack a fundamental theory. In this paper, we establish a causal inference framework, which not only unravels the whys of previous methods, but also derives a new principled solution. Specifically, our theory shows that the SGD momentum is essentially a confounder in long-tailed classification. On one hand, it has a harmful causal effect that misleads the tail prediction biased towards the head. On the other hand, its induced mediation also benefits the representation learning and head prediction. Our framework elegantly disentangles the paradoxical effects of the momentum, by pursuing the direct causal effect caused by an input sample. In particular, we use causal intervention in training, and counterfactual reasoning in inference, to remove the "bad" while keep the "good". We achieve new state-of-the-arts on three long-tailed visual recognition benchmarks: Long-tailed CIFAR-10/-100, ImageNet-LT for image classification and LVIS for instance segmentation.
Pixelwise semantic image labeling is an important, yet challenging, task with many applications. Typical approaches to tackle this problem involve either the training of deep networks on vast amounts of images to directly infer the labels or the use of probabilistic graphical models to jointly model the dependencies of the input (i.e. images) and output (i.e. labels). Yet, the former approaches do not capture the structure of the output labels, which is crucial for the performance of dense labeling, and the latter rely on carefully hand-designed priors that require costly parameter tuning via optimization techniques, which in turn leads to long inference times. To alleviate these restrictions, we explore how to arrive at dense semantic pixel labels given both the input image and an initial estimate of the output labels. We propose a parallel architecture that: 1) exploits the context information through a LabelPropagation network to propagate correct labels from nearby pixels to improve the object boundaries, 2) uses a LabelReplacement network to directly replace possibly erroneous, initial labels with new ones, and 3) combines the different intermediate results via a Fusion network to obtain the final per-pixel label. We experimentally validate our approach on two different datasets for the semantic segmentation and face parsing tasks respectively, where we show improvements over the state-of-the-art. We also provide both a quantitative and qualitative analysis of the generated results.
Face morphing attacks have raised critical concerns as they demonstrate a new vulnerability of Face Recognition Systems (FRS), which are widely deployed in border control applications. The face morphing process uses the images from multiple data subjects and performs an image blending operation to generate a morphed image of high quality. The generated morphed image exhibits similar visual characteristics corresponding to the biometric characteristics of the data subjects that contributed to the composite image and thus making it difficult for both humans and FRS, to detect such attacks. In this paper, we report a systematic investigation on the vulnerability of the Commercial-Off-The-Shelf (COTS) FRS when morphed images under the influence of ageing are presented. To this extent, we have introduced a new morphed face dataset with ageing derived from the publicly available MORPH II face dataset, which we refer to as MorphAge dataset. The dataset has two bins based on age intervals, the first bin - MorphAge-I dataset has 1002 unique data subjects with the age variation of 1 year to 2 years while the MorphAge-II dataset consists of 516 data subjects whose age intervals are from 2 years to 5 years. To effectively evaluate the vulnerability for morphing attacks, we also introduce a new evaluation metric, namely the Fully Mated Morphed Presentation Match Rate (FMMPMR), to quantify the vulnerability effectively in a realistic scenario. Extensive experiments are carried out by using two different COTS FRS (COTS I - Cognitec and COTS II - Neurotechnology) to quantify the vulnerability with ageing. Further, we also evaluate five different Morph Attack Detection (MAD) techniques to benchmark their detection performance with ageing.
A depth map can be represented by a set of learned bases and can be efficiently solved in a closed form solution. However, one issue with this method is that it may create artifacts when colour boundaries are inconsistent with depth boundaries. In fact, this is very common in a natural image. To address this issue, we enforce a more strict model in depth recovery: a piece-wise planar model. More specifically, we represent the desired depth map as a collection of 3D planar and the reconstruction problem is formulated as the optimization of planar parameters. Such a problem can be formulated as a continuous CRF optimization problem and can be solved through particle based method (MP-PBP) \cite{Yamaguchi14}. Extensive experimental evaluations on the KITTI visual odometry dataset show that our proposed methods own high resistance to false object boundaries and can generate useful and visually pleasant 3D point clouds.
Preparation of high-quality datasets for the urban scene understanding is a labor-intensive task, especially, for datasets designed for the autonomous driving applications. The application of the coarse ground truth (GT) annotations of these datasets without detriment to the accuracy of semantic image segmentation (by the mean intersection over union - mIoU) could simplify and speedup the dataset preparation and model fine tuning before its practical application. Here the results of the comparative analysis for semantic segmentation accuracy obtained by PSPNet deep learning architecture are presented for fine and coarse annotated images from Cityscapes dataset. Two scenarios were investigated: scenario 1 - the fine GT images for training and prediction, and scenario 2 - the fine GT images for training and the coarse GT images for prediction. The obtained results demonstrated that for the most important classes the mean accuracy values of semantic image segmentation for coarse GT annotations are higher than for the fine GT ones, and the standard deviation values are vice versa. It means that for some applications some unimportant classes can be excluded and the model can be tuned further for some classes and specific regions on the coarse GT dataset without loss of the accuracy even. Moreover, this opens the perspectives to use deep neural networks for the preparation of such coarse GT datasets.
In this paper we propose a novel image encryption scheme. The proposed method is based on the chaos theory. Our cryptosystem uses the chaos theory to define a dynamic chaotic Look-Up Table (LUT) to compute the new value of the current pixel to cipher. Applying this process on each pixel of the plain image, we generate the encrypted image. The results of different experimental tests, such as Key space analysis, Information Entropy and Histogram analysis, show that the proposed encryption image scheme seems to be protected against various attacks. A comparison between the plain and encrypted image, in terms of correlation coefficient, proves that the plain image is very different from the encrypted one.
While much progress has been made in 6-DoF object pose estimation from a single RGB image, the current leading approaches heavily rely on real-annotation data. As such, they remain sensitive to severe occlusions, because covering all possible occlusions with annotated data is intractable. In this paper, we introduce an approach to robustly and accurately estimate the 6-DoF pose in challenging conditions and without using any real pose annotations. To this end, we leverage the intuition that the poses predicted by a network from an image and from its counterpart synthetically altered to mimic occlusion should be consistent, and translate this to a self-supervised loss function. Our experiments on LINEMOD, Occluded-LINEMOD, YCB and new Randomization LINEMOD dataset evidence the robustness of our approach. We achieve state of the art performance on LINEMOD, and OccludedLINEMOD in without real-pose setting, even outperforming methods that rely on real annotations during training on Occluded-LINEMOD.