We present a stereo-matching method for depth estimation from high-resolution images using visual hulls as priors, and a memory-efficient technique for the correlation computation. Our method uses object masks extracted from supplementary views of the scene to guide the disparity estimation, effectively reducing the search space for matches. This approach is specifically tailored to stereo rigs in volumetric capture systems, where an accurate depth plays a key role in the downstream reconstruction task. To enable training and regression at high resolutions targeted by recent systems, our approach extends a sparse correlation computation into a hybrid sparse-dense scheme suitable for application in leading recurrent network architectures. We evaluate the performance-efficiency tradeoff of our method compared to state-of-the-art approaches and demonstrate the efficacy of the visual hull guidance. In addition, we propose a training scheme for a further reduction of memory requirements during optimization, facilitating training on high-resolution data.

Abstract Dynamic surface reconstruction of objects from point cloud sequences is a challenging field in computer graphics. Existing approaches either require multiple regularization terms or extensive training data which, however, lead to compromises in reconstruction accuracy as well as over‐smoothing or poor generalization to unseen objects and motions. To address these limitations, we introduce Preconditioned Deformation Grids , a novel technique for estimating coherent deformation fields directly from unstructured point cloud sequences without requiring or forming explicit correspondences. Key to our approach is the use of multi‐resolution voxel grids that capture the overall motion at varying spatial scales, enabling a more flexible deformation representation. In conjunction with incorporating grid‐based Sobolev preconditioning into gradient‐based optimization, we show that applying a Chamfer loss between the input point clouds as well as to an evolving template mesh is sufficient to obtain accurate deformations. To ensure temporal consistency along the object surface, we include a weak isometry loss on mesh edges which complements the main objective without constraining deformation fidelity. Extensive evaluations demonstrate that our method achieves superior results, particularly for long sequences, compared to state‐of‐the‐art techniques.

Permutation matrices play a key role in matching and assignment problems across the fields, especially in computer vision and robotics. However, memory for explicitly representing permutation matrices grows quadratically with the size of the problem, prohibiting large problem instances. In this work, we propose to tackle the curse of dimensionality of large permutation matrices by approximating them using low-rank matrix factorization, followed by a nonlinearity. To this end, we rely on the Kissing number theory to infer the minimal rank required for representing a permutation matrix of a given size, which is significantly smaller than the problem size. This leads to a drastic reduction in computation and memory costs, e.g., up to $3$ orders of magnitude less memory for a problem of size $n=20000$, represented using $8.4\times10^5$ elements in two small matrices instead of using a single huge matrix with $4\times 10^8$ elements. The proposed representation allows for accurate representations of large permutation matrices, which in turn enables handling large problems that would have been infeasible otherwise. We demonstrate the applicability and merits of the proposed approach through a series of experiments on a range of problems that involve predicting permutation matrices, from linear and quadratic assignment to shape matching problems.

Over the past decade, deep neural networks have been shown to perform extremely well on a variety of image reconstruction tasks. Such networks do, however, fail to provide guarantees about these predictions, making them difficult to use in safety-critical applications. Recent works addressed this problem by combining model-and learning-based approaches, e.g., by forcing networks to iteratively minimize a model-based cost function via the prediction of suitable descent directions. While previous approaches were limited to continuously differentiable cost functions, this paper discusses a way to remove the restriction of differentiability. We propose to use the Moreau-Yosida regularization of such costs to make the framework of energy dissipating networks applicable. We demonstrate our framework on two exemplary applications, i.e., safeguarding energy dissipating denoising networks to the expected distribution of the noise as well as enforcing binary constraints on bar-code deblurring networks to improve their respective performances.

Single image segmentation based on scribbles is an important technique in several applications, e.g. for image editing software. In this paper, we investigate the scope of single image segmentation solely given the image and scribble information using both convolutional neural networks as well as classical model-based methods, and present three main findings: 1) Despite the success of deep learning in the semantic analysis of images, networks fail to outperform model-based approaches in the case of learning on a single image only. Even using a pretrained network for transfer learning does not yield faithful segmentations. 2) The best way to utilize an annotated data set is by exploiting a model-based approach that combines semantic features of a pretrained network with the RGB information, and 3) allowing the networks prediction to change spatially and additionally enforce this variation to be smooth via a gradient-based regularization term on the loss (double backpropagation) is the most successful strategy for pure single image learning-based segmentation.

The idea of federated learning is to collaboratively train a neural network on a server. Each user receives the current weights of the network and in turns sends parameter updates (gradients) based on local data. This protocol has been designed not only to train neural networks data-efficiently, but also to provide privacy benefits for users, as their input data remains on device and only parameter gradients are shared. But how secure is sharing parameter gradients? Previous attacks have provided a false sense of security, by succeeding only in contrived settings - even for a single image. However, by exploiting a magnitude-invariant loss along with optimization strategies based on adversarial attacks, we show that is is actually possible to faithfully reconstruct images at high resolution from the knowledge of their parameter gradients, and demonstrate that such a break of privacy is possible even for trained deep networks. We analyze the effects of architecture as well as parameters on the difficulty of reconstructing an input image and prove that any input to a fully connected layer can be reconstructed analytically independent of the remaining architecture. Finally we discuss settings encountered in practice and show that even averaging gradients over several iterations or several images does not protect the user's privacy in federated learning applications.