In this study, we develop a method for multi-task manifold learning. The method aims to improve the performance of manifold learning for multiple tasks, particularly when each task has a small number of samples. Furthermore, the method also aims to generate new samples for new tasks, in addition to new samples for existing tasks. In the proposed method, we use two different types of information transfer: instance transfer and model transfer. For instance transfer, datasets are merged among similar tasks, whereas for model transfer, the manifold models are averaged among similar tasks. For this purpose, the proposed method consists of a set of generative manifold models corresponding to the tasks, which are integrated into a general model of a fiber bundle. We applied the proposed method to artificial datasets and face image sets, and the results showed that the method was able to estimate the manifolds, even for a tiny number of samples.
We explore different design choices for injecting noise into generative adversarial networks (GANs) with the goal of disentangling the latent space. Instead of traditional approaches, we propose feeding multiple noise codes through separate fully-connected layers respectively. The aim is restricting the influence of each noise code to specific parts of the generated image. We show that disentanglement in the first layer of the generator network leads to disentanglement in the generated image. Through a grid-based structure, we achieve several aspects of disentanglement without complicating the network architecture and without requiring labels. We achieve spatial disentanglement, scale-space disentanglement, and disentanglement of the foreground object from the background style allowing fine-grained control over the generated images. Examples include changing facial expressions in face images, changing beak length in bird images, and changing car dimensions in car images. This empirically leads to better disentanglement scores than state-of-the-art methods on the FFHQ dataset.
We study the effect of adversarial perturbations of images on deep stereo matching networks for the disparity estimation task. We present a method to craft a single set of perturbations that, when added to any stereo image pair in a dataset, can fool a stereo network to significantly alter the perceived scene geometry. Our perturbation images are "universal" in that they not only corrupt estimates of the network on the dataset they are optimized for, but also generalize to stereo networks with different architectures across different datasets. We evaluate our approach on multiple public benchmark datasets and show that our perturbations can increase D1-error (akin to fooling rate) of state-of-the-art stereo networks from 1% to as much as 87%. We investigate the effect of perturbations on the estimated scene geometry and identify object classes that are most vulnerable. Our analysis on the activations of registered points between left and right images led us to find that certain architectural components, i.e. deformable convolution and explicit matching, can increase robustness against adversaries. We demonstrate that by simply designing networks with such components, one can reduce the effect of adversaries by up to 60.5%, which rivals the robustness of networks fine-tuned with costly adversarial data augmentation.
Gradient inversion attack (or input recovery from gradient) is an emerging threat to the security and privacy preservation of Federated learning, whereby malicious eavesdroppers or participants in the protocol can recover (partially) the clients' private data. This paper evaluates existing attacks and defenses. We find that some attacks make strong assumptions about the setup. Relaxing such assumptions can substantially weaken these attacks. We then evaluate the benefits of three proposed defense mechanisms against gradient inversion attacks. We show the trade-offs of privacy leakage and data utility of these defense methods, and find that combining them in an appropriate manner makes the attack less effective, even under the original strong assumptions. We also estimate the computation cost of end-to-end recovery of a single image under each evaluated defense. Our findings suggest that the state-of-the-art attacks can currently be defended against with minor data utility loss, as summarized in a list of potential strategies. Our code is available at: https://github.com/Princeton-SysML/GradAttack.
In medical imaging it is common practice to acquire a wide range of modalities (MRI, CT, PET, etc.), to highlight different structures or pathologies. As patient movement between scans or scanning session is unavoidable, registration is often an essential step before any subsequent image analysis. In this paper, we introduce a cost function based on joint total variation for such multimodal image registration. This cost function has the advantage of enabling principled, groupwise alignment of multiple images, whilst being insensitive to strong intensity non-uniformities. We evaluate our algorithm on rigidly aligning both simulated and real 3D brain scans. This validation shows robustness to strong intensity non-uniformities and low registration errors for CT/PET to MRI alignment. Our implementation is publicly available at https://github.com/brudfors/coregistration-njtv.
360{\deg} cameras can capture complete environments in a single shot, which makes 360{\deg} imagery alluring in many computer vision tasks. However, monocular depth estimation remains a challenge for 360{\deg} data, particularly for high resolutions like 2K (2048$\times$1024) that are important for novel-view synthesis and virtual reality applications. Current CNN-based methods do not support such high resolutions due to limited GPU memory. In this work, we propose a flexible framework for monocular depth estimation from high-resolution 360{\deg} images using tangent images. We project the 360{\deg} input image onto a set of tangent planes that produce perspective views, which are suitable for the latest, most accurate state-of-the-art perspective monocular depth estimators. We recombine the individual depth estimates using deformable multi-scale alignment followed by gradient-domain blending to improve the consistency of disparity estimates. The result is a dense, high-resolution 360{\deg} depth map with a high level of detail, also for outdoor scenes which are not supported by existing methods.
Determining the poverty levels of various regions throughout the world is crucial in identifying interventions for poverty reduction initiatives and directing resources fairly. However, reliable data on global economic livelihoods is hard to come by, especially for areas in the developing world, hampering efforts to both deploy services and monitor/evaluate progress. This is largely due to the fact that this data is obtained from traditional door-to-door surveys, which are time consuming and expensive. Overhead satellite imagery contain characteristics that make it possible to estimate the region's poverty level. In this work, I develop deep learning computer vision methods that can predict a region's poverty level from an overhead satellite image. I experiment with both daytime and nighttime imagery. Furthermore, because data limitations are often the barrier to entry in poverty prediction from satellite imagery, I explore the impact that data quantity and data augmentation have on the representational power and overall accuracy of the networks. Lastly, to evaluate the robustness of the networks, I evaluate them on data from continents that were absent in the development set.
Understanding document images (e.g., invoices) has been an important research topic and has many applications in document processing automation. Through the latest advances in deep learning-based Optical Character Recognition (OCR), current Visual Document Understanding (VDU) systems have come to be designed based on OCR. Although such OCR-based approach promise reasonable performance, they suffer from critical problems induced by the OCR, e.g., (1) expensive computational costs and (2) performance degradation due to the OCR error propagation. In this paper, we propose a novel VDU model that is end-to-end trainable without underpinning OCR framework. To this end, we propose a new task and a synthetic document image generator to pre-train the model to mitigate the dependencies on large-scale real document images. Our approach achieves state-of-the-art performance on various document understanding tasks in public benchmark datasets and private industrial service datasets. Through extensive experiments and analysis, we demonstrate the effectiveness of the proposed model especially with consideration for a real-world application.
Facial sketches drawn by artists are widely used for visual identification applications and mostly by law enforcement agencies, but the quality of these sketches depend on the ability of the artist to clearly replicate all the key facial features that could aid in capturing the true identity of a subject. Recent works have attempted to synthesize these sketches into plausible visual images to improve visual recognition and identification. However, synthesizing photo-realistic images from sketches proves to be an even more challenging task, especially for sensitive applications such as suspect identification. In this work, we propose a novel approach that adopts a generative adversarial network that synthesizes a single sketch into multiple synthetic images with unique attributes like hair color, sex, etc. We incorporate a hybrid discriminator which performs attribute classification of multiple target attributes, a quality guided encoder that minimizes the perceptual dissimilarity of the latent space embedding of the synthesized and real image at different layers in the network and an identity preserving network that maintains the identity of the synthesised image throughout the training process. Our approach is aimed at improving the visual appeal of the synthesised images while incorporating multiple attribute assignment to the generator without compromising the identity of the synthesised image. We synthesised sketches using XDOG filter for the CelebA, WVU Multi-modal and CelebA-HQ datasets and from an auxiliary generator trained on sketches from CUHK, IIT-D and FERET datasets. Our results are impressive compared to current state of the art.
We propose a novel approach to translate unpaired contrast computed tomography (CT) scans to non-contrast CT scans and the other way around. Solving this task has two important applications: (i) to automatically generate contrast CT scans for patients for whom injecting contrast substance is not an option, and (ii) to enhance alignment between contrast and non-contrast CT by reducing the differences induced by the contrast substance before registration. Our approach is based on cycle-consistent generative adversarial convolutional transformers, for short, CyTran. Our neural model can be trained on unpaired images, due to the integration of a cycle-consistency loss. To deal with high-resolution images, we design a hybrid architecture based on convolutional and multi-head attention layers. In addition, we introduce a novel data set, Coltea-Lung-CT-100W, containing 3D triphasic lung CT scans (with a total of 37,290 images) collected from 100 female patients. Each scan contains three phases (non-contrast, early portal venous, and late arterial), allowing us to perform experiments to compare our novel approach with state-of-the-art methods for image style transfer. Our empirical results show that CyTran outperforms all competing methods. Moreover, we show that CyTran can be employed as a preliminary step to improve a state-of-the-art medical image alignment method. We release our novel model and data set as open source at: https://github.com/ristea/cycle-transformer.