Michael Moeller

Marius Bock, Hilde Kuehne, Kristof Van Laerhoven et al. · ibm-research, mit

Research has shown the complementarity of camera- and inertial-based data for modeling human activities, yet datasets with both egocentric video and inertial-based sensor data remain scarce. In this paper, we introduce WEAR, an outdoor sports dataset for both vision- and inertial-based human activity recognition (HAR). Data from 22 participants performing a total of 18 different workout activities was collected with synchronized inertial (acceleration) and camera (egocentric video) data recorded at 11 different outside locations. WEAR provides a challenging prediction scenario in changing outdoor environments using a sensor placement, in line with recent trends in real-world applications. Benchmark results show that through our sensor placement, each modality interestingly offers complementary strengths and weaknesses in their prediction performance. Further, in light of the recent success of single-stage Temporal Action Localization (TAL) models, we demonstrate their versatility of not only being trained using visual data, but also using raw inertial data and being capable to fuse both modalities by means of simple concatenation. The dataset and code to reproduce experiments is publicly available via: mariusbock.github.io/wear/.

Martin Burger, Guy Gilboa, Michael Moeller et al.

This paper discusses the use of absolutely one-homogeneous regularization functionals in a variational, scale space, and inverse scale space setting to define a nonlinear spectral decomposition of input data. We present several theoretical results that explain the relation between the different definitions. Additionally, results on the orthogonality of the decomposition, a Parseval-type identity and the notion of generalized (nonlinear) eigenvectors closely link our nonlinear multiscale decompositions to the well-known linear filtering theory. Numerical results are used to illustrate our findings.

Marcel Seelbach Benkner, Maximilian Krahn, Edith Tretschk et al.

Modern quantum annealers can find high-quality solutions to combinatorial optimisation objectives given as quadratic unconstrained binary optimisation (QUBO) problems. Unfortunately, obtaining suitable QUBO forms in computer vision remains challenging and currently requires problem-specific analytical derivations. Moreover, such explicit formulations impose tangible constraints on solution encodings. In stark contrast to prior work, this paper proposes to learn QUBO forms from data through gradient backpropagation instead of deriving them. As a result, the solution encodings can be chosen flexibly and compactly. Furthermore, our methodology is general and virtually independent of the specifics of the target problem type. We demonstrate the advantages of learnt QUBOs on the diverse problem types of graph matching, 2D point cloud alignment and 3D rotation estimation. Our results are competitive with the previous quantum state of the art while requiring much fewer logical and physical qubits, enabling our method to scale to larger problems. The code and the new dataset will be open-sourced.

Lukas Koestler, Daniel Grittner, Michael Moeller et al.

Neural fields have gained significant attention in the computer vision community due to their excellent performance in novel view synthesis, geometry reconstruction, and generative modeling. Some of their advantages are a sound theoretic foundation and an easy implementation in current deep learning frameworks. While neural fields have been applied to signals on manifolds, e.g., for texture reconstruction, their representation has been limited to extrinsically embedding the shape into Euclidean space. The extrinsic embedding ignores known intrinsic manifold properties and is inflexible wrt. transfer of the learned function. To overcome these limitations, this work introduces intrinsic neural fields, a novel and versatile representation for neural fields on manifolds. Intrinsic neural fields combine the advantages of neural fields with the spectral properties of the Laplace-Beltrami operator. We show theoretically that intrinsic neural fields inherit many desirable properties of the extrinsic neural field framework but exhibit additional intrinsic qualities, like isometry invariance. In experiments, we show intrinsic neural fields can reconstruct high-fidelity textures from images with state-of-the-art quality and are robust to the discretization of the underlying manifold. We demonstrate the versatility of intrinsic neural fields by tackling various applications: texture transfer between deformed shapes & different shapes, texture reconstruction from real-world images with view dependence, and discretization-agnostic learning on meshes and point clouds.

Harshil Bhatia, Edith Tretschk, Zorah Lähner et al.

Jointly matching multiple, non-rigidly deformed 3D shapes is a challenging, $\mathcal{NP}$-hard problem. A perfect matching is necessarily cycle-consistent: Following the pairwise point correspondences along several shapes must end up at the starting vertex of the original shape. Unfortunately, existing quantum shape-matching methods do not support multiple shapes and even less cycle consistency. This paper addresses the open challenges and introduces the first quantum-hybrid approach for 3D shape multi-matching; in addition, it is also cycle-consistent. Its iterative formulation is admissible to modern adiabatic quantum hardware and scales linearly with the total number of input shapes. Both these characteristics are achieved by reducing the $N$-shape case to a sequence of three-shape matchings, the derivation of which is our main technical contribution. Thanks to quantum annealing, high-quality solutions with low energy are retrieved for the intermediate $\mathcal{NP}$-hard objectives. On benchmark datasets, the proposed approach significantly outperforms extensions to multi-shape matching of a previous quantum-hybrid two-shape matching method and is on-par with classical multi-matching methods.

Maolin Gao, Paul Roetzer, Marvin Eisenberger et al.

We propose a novel mixed-integer programming (MIP) formulation for generating precise sparse correspondences for highly non-rigid shapes. To this end, we introduce a projected Laplace-Beltrami operator (PLBO) which combines intrinsic and extrinsic geometric information to measure the deformation quality induced by predicted correspondences. We integrate the PLBO, together with an orientation-aware regulariser, into a novel MIP formulation that can be solved to global optimality for many practical problems. In contrast to previous methods, our approach is provably invariant to rigid transformations and global scaling, initialisation-free, has optimality guarantees, and scales to high resolution meshes with (empirically observed) linear time. We show state-of-the-art results for sparse non-rigid matching on several challenging 3D datasets, including data with inconsistent meshing, as well as applications in mesh-to-point-cloud matching.

Samira Kabri, Alexander Auras, Danilo Riccio et al.

The reconstruction of images from their corresponding noisy Radon transform is a typical example of an ill-posed linear inverse problem as arising in the application of computerized tomography (CT). As the (naive) solution does not depend on the measured data continuously, regularization is needed to re-establish a continuous dependence. In this work, we investigate simple, but yet still provably convergent approaches to learning linear regularization methods from data. More specifically, we analyze two approaches: One generic linear regularization that learns how to manipulate the singular values of the linear operator in an extension of our previous work, and one tailored approach in the Fourier domain that is specific to CT-reconstruction. We prove that such approaches become convergent regularization methods as well as the fact that the reconstructions they provide are typically much smoother than the training data they were trained on. Finally, we compare the spectral as well as the Fourier-based approaches for CT-reconstruction numerically, discuss their advantages and disadvantages and investigate the effect of discretization errors at different resolutions.

Marius Bock, Michael Moeller, Kristof Van Laerhoven

As of today, state-of-the-art activity recognition from wearable sensors relies on algorithms being trained to classify fixed windows of data. In contrast, video-based Human Activity Recognition, known as Temporal Action Localization (TAL), has followed a segment-based prediction approach, localizing activity segments in a timeline of arbitrary length. This paper is the first to systematically demonstrate the applicability of state-of-the-art TAL models for both offline and near-online Human Activity Recognition (HAR) using raw inertial data as well as pre-extracted latent features as input. Offline prediction results show that TAL models are able to outperform popular inertial models on a multitude of HAR benchmark datasets, with improvements reaching as much as 26% in F1-score. We show that by analyzing timelines as a whole, TAL models can produce more coherent segments and achieve higher NULL-class accuracy across all datasets. We demonstrate that TAL is less suited for the immediate classification of small-sized windows of data, yet offers an interesting perspective on inertial-based HAR -- alleviating the need for fixed-size windows and enabling algorithms to recognize activities of arbitrary length. With design choices and training concepts yet to be explored, we argue that TAL architectures could be of significant value to the inertial-based HAR community. The code and data download to reproduce experiments is publicly available via github.com/mariusbock/tal_for_har.

Hendrik Sommerhoff, Shashank Agnihotri, Mohamed Saleh et al.

The success of deep learning is frequently described as the ability to train all parameters of a network on a specific application in an end-to-end fashion. Yet, several design choices on the camera level, including the pixel layout of the sensor, are considered as pre-defined and fixed, and high resolution, regular pixel layouts are considered to be the most generic ones in computer vision and graphics, treating all regions of an image as equally important. While several works have considered non-uniform, \eg, hexagonal or foveated, pixel layouts in hardware and image processing, the layout has not been integrated into the end-to-end learning paradigm so far. In this work, we present the first truly end-to-end trained imaging pipeline that optimizes the size and distribution of pixels on the imaging sensor jointly with the parameters of a given neural network on a specific task. We derive an analytic, differentiable approach for the sensor layout parameterization that allows for task-specific, local varying pixel resolutions. We present two pixel layout parameterization functions: rectangular and curvilinear grid shapes that retain a regular topology. We provide a drop-in module that approximates sensor simulation given existing high-resolution images to directly connect our method with existing deep learning models. We show that network predictions benefit from learnable pixel layouts for two different downstream tasks, classification and semantic segmentation.

Marius Bock, Kristof Van Laerhoven, Michael Moeller

As wearable-based data annotation remains, to date, a tedious, time-consuming task requiring researchers to dedicate substantial time, benchmark datasets within the field of Human Activity Recognition in lack richness and size compared to datasets available within related fields. Recently, vision foundation models such as CLIP have gained significant attention, helping the vision community advance in finding robust, generalizable feature representations. With the majority of researchers within the wearable community relying on vision modalities to overcome the limited expressiveness of wearable data and accurately label their to-be-released benchmark datasets offline, we propose a novel, clustering-based annotation pipeline to significantly reduce the amount of data that needs to be annotated by a human annotator. We show that using our approach, the annotation of centroid clips suffices to achieve average labelling accuracies close to 90% across three publicly available HAR benchmark datasets. Using the weakly annotated datasets, we further demonstrate that we can match the accuracy scores of fully-supervised deep learning classifiers across all three benchmark datasets. Code as well as supplementary figures and results are publicly downloadable via github.com/mariusbock/weak_har.

Kanchana Vaishnavi Gandikota, Jonas Geiping, Zorah Lähner et al.

Many applications require robustness, or ideally invariance, of neural networks to certain transformations of input data. Most commonly, this requirement is addressed by training data augmentation, using adversarial training, or defining network architectures that include the desired invariance by design. In this work, we propose a method to make network architectures provably invariant with respect to group actions by choosing one element from a (possibly continuous) orbit based on a fixed criterion. In a nutshell, we intend to 'undo' any possible transformation before feeding the data into the actual network. Further, we empirically analyze the properties of different approaches which incorporate invariance via training or architecture, and demonstrate the advantages of our method in terms of robustness and computational efficiency. In particular, we investigate the robustness with respect to rotations of images (which can hold up to discretization artifacts) as well as the provable orientation and scaling invariance of 3D point cloud classification.

Jovita Lukasik, Michael Moeller, Margret Keuper

Zero-cost proxies are nowadays frequently studied and used to search for neural architectures. They show an impressive ability to predict the performance of architectures by making use of their untrained weights. These techniques allow for immense search speed-ups. So far the joint search for well-performing and robust architectures has received much less attention in the field of NAS. Therefore, the main focus of zero-cost proxies is the clean accuracy of architectures, whereas the model robustness should play an evenly important part. In this paper, we analyze the ability of common zero-cost proxies to serve as performance predictors for robustness in the popular NAS-Bench-201 search space. We are interested in the single prediction task for robustness and the joint multi-objective of clean and robust accuracy. We further analyze the feature importance of the proxies and show that predicting the robustness makes the prediction task from existing zero-cost proxies more challenging. As a result, the joint consideration of several proxies becomes necessary to predict a model's robustness while the clean accuracy can be regressed from a single such feature.

Kanchana Vaishnavi Gandikota, Paramanand Chandramouli, Michael Moeller

Recent approaches employ deep learning-based solutions for the recovery of a sharp image from its blurry observation. This paper introduces adversarial attacks against deep learning-based image deblurring methods and evaluates the robustness of these neural networks to untargeted and targeted attacks. We demonstrate that imperceptible distortion can significantly degrade the performance of state-of-the-art deblurring networks, even producing drastically different content in the output, indicating the strong need to include adversarially robust training not only in classification but also for image recovery.

Mishal Fatima, Shashank Agnihotri, Kanchana Vaishnavi Gandikota et al.

Most vision models are trained on RGB images processed through ISP pipelines optimized for human perception, which can discard sensor-level information useful for machine reasoning. RAW images preserve unprocessed scene data, enabling models to leverage richer cues for both object detection and object description, capturing fine-grained details, spatial relationships, and contextual information often lost in processed images. To support research in this domain, we introduce RAWDet-7, a large-scale dataset of ~25k training and 7.6k test RAW images collected across diverse cameras, lighting conditions, and environments, densely annotated for seven object categories following MS-COCO and LVIS conventions. In addition, we provide object-level descriptions derived from the corresponding high-resolution sRGB images, facilitating the study of object-level information preservation under RAW image processing and low-bit quantization. The dataset allows evaluation under simulated 4-bit, 6-bit, and 8-bit quantization, reflecting realistic sensor constraints, and provides a benchmark for studying detection performance, description quality & detail, and generalization in low-bit RAW image processing. Dataset & code upon acceptance.

Natacha Kuete Meli, Jovita Lukasik, Vladislav Golyanik et al.

We introduce layered Quantum Architecture Search (layered-QAS), a strategy inspired by classical network morphism that designs Parametrised Quantum Circuit (PQC) architectures by progressively growing and adapting them. PQCs offer strong expressiveness with relatively few parameters, yet they lack standard architectural layers (e.g., convolution, attention) that encode inductive biases for a given learning task. To assess the effectiveness of our method, we focus on 3D point cloud classification as a challenging yet highly structured problem. Whereas prior work on this task has used PQCs only as feature extractors for classical classifiers, our approach uses the PQC as the main building block of the classification model. Simulations show that our layered-QAS mitigates barren plateau, outperforms quantum-adapted local and evolutionary QAS baselines, and achieves state-of-the-art results among PQC-based methods on the ModelNet dataset.

Sonali Godavarthy, Matthias Neuwirth-Trapp, Tim-Felix Faasch et al.

Recent text-to-image models produce high-quality images, yet text ambiguity hinders precise control when specific styles or objects are required. There have been a number of recent works dealing with learning and composing multiple objects and patterns. However, current work focuses almost entirely on image content, overlooking imaging factors such as camera lens, sensor types, imaging viewpoints, and scenes' domain characteristics. We introduce this new challenge as Imaging Factor Disentanglement and show limitations of current approaches in the regime. We, therefore, propose the new method Multi-factor disentanglement through Textual Inversion (MULTI). It consists of two stages: in the first stage, we learn general factors, and in the second stage, we extract dataset-specific ones. This setup enables the extension of existing datasets and novel factor combinations, thereby reducing distribution gaps. It further supports modifications of specific factors and image-to-image generation via ControlNets. The evaluation on our new DF-RICO benchmark demonstrates the effectiveness of MULTI and highlights the importance of Factor Disentanglement as a new direction of research.

Shiwen Qin, Alexander Auras, Shay B. Cohen et al.

Neural architecture search (NAS) automates the design process of high-performing architectures, but remains bottlenecked by expensive performance evaluation. Most existing studies that achieve faster evaluation are mostly tied to cell-based search spaces and graph encodings tailored to those individual search spaces, limiting their flexibility and scalability when applied to more expressive search spaces. In this work, we aim to close the gap of individual search space restrictions and search space dependent network representations. We present ONNX-Bench, a benchmark consisting of a collection of neural networks in a unified format based on ONNX files. ONNX-Bench includes all open-source NAS-bench-based neural networks, resulting in a total size of more than 600k {architecture, accuracy} pairs. This benchmark allows creating a shared neural network representation, ONNX-Net, able to represent any neural architecture using natural language descriptions acting as an input to a performance predictor. This text-based encoding can accommodate arbitrary layer types, operation parameters, and heterogeneous topologies, enabling a single surrogate to generalise across all neural architectures rather than being confined to cell-based search spaces. Experiments show strong zero-shot performance across disparate search spaces using only a small amount of pretraining samples, enabling the unprecedented ability to evaluate any neural network architecture instantly.

Mishal Fatima, Shashank Agnihotri, Marius Bock et al.

Most pattern recognition models are developed on pre-proce\-ssed data. In computer vision, for instance, RGB images processed through image signal processing (ISP) pipelines designed to cater to human perception are the most frequent input to image analysis networks. However, many modern vision tasks operate without a human in the loop, raising the question of whether such pre-processing is optimal for automated analysis. Similarly, human activity recognition (HAR) on body-worn sensor data commonly takes normalized floating-point data arising from a high-bit analog-to-digital converter (ADC) as an input, despite such an approach being highly inefficient in terms of data transmission, significantly affecting the battery life of wearable devices. In this work, we target low-bandwidth and energy-constrained settings where sensors are limited to low-bit-depth capture. We propose $γ$-Quant, i.e.~the task-specific learning of a non-linear quantization for pattern recognition. We exemplify our approach on raw-image object detection as well as HAR of wearable data, and demonstrate that raw data with a learnable quantization using as few as 4-bits can perform on par with the use of raw 12-bit data. All code to reproduce our experiments is publicly available via https://github.com/Mishalfatima/Gamma-Quant

Marius Bock, Michael Moeller, Kristof Van Laerhoven

Despite recognized limitations in modeling long-range temporal dependencies, Human Activity Recognition (HAR) has traditionally relied on a sliding window approach to segment labeled datasets. Deep learning models like the DeepConvLSTM typically classify each window independently, thereby restricting learnable temporal context to within-window information. To address this constraint, we propose DeepConvContext, a multi-scale time series classification framework for HAR. Drawing inspiration from the vision-based Temporal Action Localization community, DeepConvContext models both intra- and inter-window temporal patterns by processing sequences of time-ordered windows. Unlike recent HAR models that incorporate attention mechanisms, DeepConvContext relies solely on LSTMs -- with ablation studies demonstrating the superior performance of LSTMs over attention-based variants for modeling inertial sensor data. Across six widely-used HAR benchmarks, DeepConvContext achieves an average 10% improvement in F1-score over the classic DeepConvLSTM, with gains of up to 21%. Code to reproduce our experiments is publicly available via github.com/mariusbock/context_har.

Marius Bock, Maximilian Hopp, Kristof Van Laerhoven et al.

While prior work has shown that Federated Learning updates can leak sensitive information, label reconstruction attacks, which aim to recover input labels from shared gradients, have not yet been examined in the context of Human Activity Recognition (HAR). Given the sensitive nature of activity labels, this study evaluates the effectiveness of state-of-the-art gradient-based label leakage attacks on HAR benchmark datasets. Our findings show that the number of activity classes, sampling strategy, and class imbalance are critical factors influencing the extent of label leakage, with reconstruction accuracies reaching well-above 90% on two benchmark datasets, even for trained models. Moreover, we find that Local Differential Privacy techniques such as gradient noise and clipping offer only limited protection, as certain attacks still reliably infer both majority and minority class labels. We conclude by offering practical recommendations for the privacy-aware deployment of federated HAR systems and identify open challenges for future research. Code to reproduce our experiments is publicly available via github.com/mariusbock/leakage_har.

Hendrik Sommerhoff, Michael Moeller

Plug-and-play (PnP) methods for solving inverse problems have recently achieved strong performance by leveraging denoising priors based on powerful generative diffusion and flow models. However, existing diffusion- and flow-based PnP methods typically rely on stochastic renoise-denoise operations, which complicate the analysis of their convergence behavior. In this work, we identify and formalize the deterministic renoise-denoise operator underlying flow-based plug-and-play methods. This perspective reveals that these methods implicitly define a deterministic operator given by the expectation of a denoiser over the latent noise distribution. Building on this insight, we propose FlowADMM, a PnP algorithm that integrates the renoise-denoise operator into the classical alternating direction method of multiplier (ADMM) framework. We establish convergence guarantees for FlowADMM under weak Lipschitz conditions on the underlying flow network, and extend the analysis to non-stationary time schedules. Empirically, FlowADMM achieves state-of-the-art performance among flow-based PnP methods on a range of inverse problems, including denoising, deblurring, super-resolution, and inpainting, while requiring fewer data consistency evaluations than prior approaches.

Kanchana Vaishnavi Gandikota, Paramanand Chandramouli, Hannah Droege et al.

Low dose computed tomography (CT) acquisition using reduced radiation or sparse angle measurements is recommended to decrease the harmful effects of X-ray radiation. Recent works successfully apply deep networks to the problem of low dose CT recovery on bench-mark datasets. However, their robustness needs a thorough evaluation before use in clinical settings. In this work, we evaluate the robustness of different deep learning approaches and classical methods for CT recovery. We show that deep networks, including model-based networks encouraging data consistency, are more susceptible to untargeted attacks. Surprisingly, we observe that data consistency is not heavily affected even for these poor quality reconstructions, motivating the need for better regularization for the networks. We demonstrate the feasibility of universal attacks and study attack transferability across different methods. We analyze robustness to attacks causing localized changes in clinically relevant regions. Both classical approaches and deep networks are affected by such attacks leading to changes in the visual appearance of localized lesions, for extremely small perturbations. As the resulting reconstructions have high data consistency with the original measurements, these localized attacks can be used to explore the solution space of the CT recovery problem.

Christina Runkel, Kanchana Vaishnavi Gandikota, Jonas Geiping et al.

Being able to reconstruct training data from the parameters of a neural network is a major privacy concern. Previous works have shown that reconstructing training data, under certain circumstances, is possible. In this work, we analyse such reconstructions empirically and propose a new formulation of the reconstruction as a solution to a bilevel optimisation problem. We demonstrate that our formulation as well as previous approaches highly depend on the initialisation of the training images $x$ to reconstruct. In particular, we show that a random initialisation of $x$ can lead to reconstructions that resemble valid training samples while not being part of the actual training dataset. Thus, our experiments on affine and one-hidden layer networks suggest that when reconstructing natural images, yet an adversary cannot identify whether reconstructed images have indeed been part of the set of training samples.

Jan Philipp Schneider, Pratik Singh Bisht, Ilya Chugunov et al.

Learning editable high-resolution scene representations for dynamic scenes is an open problem with applications across the domains from autonomous driving to creative editing - the most successful approaches today make a trade-off between editability and supporting scene complexity: neural atlases represent dynamic scenes as two deforming image layers, foreground and background, which are editable in 2D, but break down when multiple objects occlude and interact. In contrast, scene graph models make use of annotated data such as masks and bounding boxes from autonomous-driving datasets to capture complex 3D spatial relationships, but their implicit volumetric node representations are challenging to edit view-consistently. We propose Neural Atlas Graphs (NAGs), a hybrid high-resolution scene representation, where every graph node is a view-dependent neural atlas, facilitating both 2D appearance editing and 3D ordering and positioning of scene elements. Fit at test-time, NAGs achieve state-of-the-art quantitative results on the Waymo Open Dataset - by 5 dB PSNR increase compared to existing methods - and make environmental editing possible in high resolution and visual quality - creating counterfactual driving scenarios with new backgrounds and edited vehicle appearance. We find that the method also generalizes beyond driving scenes and compares favorably - by more than 7 dB in PSNR - to recent matting and video editing baselines on the DAVIS video dataset with a diverse set of human and animal-centric scenes. Project Page: https://princeton-computational-imaging.github.io/nag/

Alexander Auras, Kanchana Vaishnavi Gandikota, Hannah Droege et al.

This paper provides an overview of current approaches for solving inverse problems in imaging using variational methods and machine learning. A special focus lies on point estimators and their robustness against adversarial perturbations. In this context results of numerical experiments for a one-dimensional toy problem are provided, showing the robustness of different approaches and empirically verifying theoretical guarantees. Another focus of this review is the exploration of the subspace of data-consistent solutions through explicit guidance to satisfy specific semantic or textural properties.

Natacha Kuete Meli, Shuteng Wang, Marcel Seelbach Benkner et al.

Quantum-enhanced Computer Vision (QeCV) is a new research field at the intersection of computer vision, optimisation theory, machine learning and quantum computing. It has high potential to transform how visual signals are processed and interpreted with the help of quantum computing that leverages quantum-mechanical effects in computations inaccessible to classical (i.e. non-quantum) computers. In scenarios where existing non-quantum methods cannot find a solution in a reasonable time or compute only approximate solutions, quantum computers can provide, among others, advantages in terms of better time scalability for multiple problem classes. Parametrised quantum circuits can also become, in the long term, a considerable alternative to classical neural networks in computer vision. However, specialised and fundamentally new algorithms must be developed to enable compatibility with quantum hardware and unveil the potential of quantum computational paradigms in computer vision. This survey contributes to the existing literature on QeCV with a holistic review of this research field. It is designed as a quantum computing reference for the computer vision community, targeting computer vision students, scientists and readers with related backgrounds who want to familiarise themselves with QeCV. We provide a comprehensive introduction to QeCV, its specifics, and methodologies for formulations compatible with quantum hardware and QeCV methods, leveraging two main quantum computational paradigms, i.e. gate-based quantum computing and quantum annealing. We elaborate on the operational principles of quantum computers and the available tools to access, program and simulate them in the context of QeCV. Finally, we review existing quantum computing tools and learning materials and discuss aspects related to publishing and reviewing QeCV papers, open challenges and potential social implications.

Christina Runkel, Natacha Kuete Meli, Jovita Lukasik et al.

Compressing and pruning large machine learning models has become a critical step towards their deployment in real-world applications. Standard pruning and compression techniques are typically designed without taking the structure of the network's weights into account, limiting their effectiveness. We explore the impact of smooth regularization on neural network training and model compression. By applying nuclear norm, first- and second-order derivative penalties of the weights during training, we encourage structured smoothness while preserving predictive performance on par with non-smooth models. We find that standard pruning methods often perform better when applied to these smooth models. Building on this observation, we apply a Singular-Value-Decomposition-based compression method that exploits the underlying smooth structure and approximates the model's weight tensors by smaller low-rank tensors. Our approach enables state-of-the-art compression without any fine-tuning - reaching up to $91\%$ accuracy on a smooth ResNet-18 on CIFAR-10 with $70\%$ fewer parameters.

Navya Sonal Agarwal, Jan Philipp Schneider, Kanchana Vaishnavi Gandikota et al.

The computational imaging technique of Fourier Ptychographic Microscopy (FPM) enables high-resolution imaging with a wide field of view and can serve as an extremely valuable tool, e.g. in the classification of cells in medical applications. However, reconstructing a high-resolution image from tens or even hundreds of measurements is computationally expensive, particularly for a wide field of view. Therefore, in this paper, we investigate the idea of classifying the image content in the FPM measurements directly without performing a reconstruction step first. We show that Convolutional Neural Networks (CNN) can extract meaningful information from measurement sequences, significantly outperforming the classification on a single band-limited image (up to 12 %) while being significantly more efficient than a reconstruction of a high-resolution image. Furthermore, we demonstrate that a learned multiplexing of several raw measurements allows maintaining the classification accuracy while reducing the amount of data (and consequently also the acquisition time) significantly.

Michael Schopf-Kuester, Zorah Lähner, Michael Moeller

This work addresses the problem of \textit{shape completion}, i.e., the task of restoring incomplete shapes by predicting their missing parts. While previous works have often predicted the fractured and restored shape in one step, we approach the task by separately predicting the fractured and newly restored parts, but ensuring these predictions are interconnected. We use a decoder network motivated by related work on the prediction of signed distance functions (DeepSDF). In particular, our representation allows us to consider test-time-training, i.e., finetuning network parameters to match the given incomplete shape more accurately during inference. While previous works often have difficulties with artifacts around the fracture boundary, we demonstrate that our overfitting to the fractured parts leads to significant improvements in the restoration of eight different shape categories of the ShapeNet data set in terms of their chamfer distances.

Marius Bock, Alexander Hoelzemann, Michael Moeller et al.

Activity recognition systems that are capable of estimating human activities from wearable inertial sensors have come a long way in the past decades. Not only have state-of-the-art methods moved away from feature engineering and have fully adopted end-to-end deep learning approaches, best practices for setting up experiments, preparing datasets, and validating activity recognition approaches have similarly evolved. This tutorial was first held at the 2021 ACM International Symposium on Wearable Computers (ISWC'21) and International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp'21). The tutorial, after a short introduction in the research field of activity recognition, provides a hands-on and interactive walk-through of the most important steps in the data pipeline for the deep learning of human activities. All presentation slides shown during the tutorial, which also contain links to all code exercises, as well as the link of the GitHub page of the tutorial can be found on: https://mariusbock.github.io/dl-for-har

Jonas Geiping, Micah Goldblum, Phillip E. Pope et al.

It is widely believed that the implicit regularization of SGD is fundamental to the impressive generalization behavior we observe in neural networks. In this work, we demonstrate that non-stochastic full-batch training can achieve comparably strong performance to SGD on CIFAR-10 using modern architectures. To this end, we show that the implicit regularization of SGD can be completely replaced with explicit regularization even when comparing against a strong and well-researched baseline. Our observations indicate that the perceived difficulty of full-batch training may be the result of its optimization properties and the disproportionate time and effort spent by the ML community tuning optimizers and hyperparameters for small-batch training.

Jonas Geiping, Jovita Lukasik, Margret Keuper et al.

Differentiable architecture search (DAS) is a widely researched tool for the discovery of novel architectures, due to its promising results for image classification. The main benefit of DAS is the effectiveness achieved through the weight-sharing one-shot paradigm, which allows efficient architecture search. In this work, we investigate DAS in a systematic case study of inverse problems, which allows us to analyze these potential benefits in a controlled manner. We demonstrate that the success of DAS can be extended from image classification to signal reconstruction, in principle. However, our experiments also expose three fundamental difficulties in the evaluation of DAS-based methods in inverse problems: First, the results show a large variance in all test cases. Second, the final performance is strongly dependent on the hyperparameters of the optimizer. And third, the performance of the weight-sharing architecture used during training does not reflect the final performance of the found architecture well. While the results on image reconstruction confirm the potential of the DAS paradigm, they challenge the common understanding of DAS as a one-shot method.

Marius Bock, Alexander Hoelzemann, Michael Moeller et al.

Recent studies in Human Activity Recognition (HAR) have shown that Deep Learning methods are able to outperform classical Machine Learning algorithms. One popular Deep Learning architecture in HAR is the DeepConvLSTM. In this paper we propose to alter the DeepConvLSTM architecture to employ a 1-layered instead of a 2-layered LSTM. We validate our architecture change on 5 publicly available HAR datasets by comparing the predictive performance with and without the change employing varying hidden units within the LSTM layer(s). Results show that across all datasets, our architecture consistently improves on the original one: Recognition performance increases up to 11.7% for the F1-score, and our architecture significantly decreases the amount of learnable parameters. This improvement over DeepConvLSTM decreases training time by as much as 48%. Our results stand in contrast to the belief that one needs at least a 2-layered LSTM when dealing with sequential data. Based on our results we argue that said claim might not be applicable to sensor-based HAR.

Hartmut Bauermeister, Emanuel Laude, Thomas Möllenhoff et al.

Dual decomposition approaches in nonconvex optimization may suffer from a duality gap. This poses a challenge when applying them directly to nonconvex problems such as MAP-inference in a Markov random field (MRF) with continuous state spaces. To eliminate such gaps, this paper considers a reformulation of the original nonconvex task in the space of measures. This infinite-dimensional reformulation is then approximated by a semi-infinite one, which is obtained via a piecewise polynomial discretization in the dual. We provide a geometric intuition behind the primal problem induced by the dual discretization and draw connections to optimization over moment spaces. In contrast to existing discretizations which suffer from a grid bias, we show that a piecewise polynomial discretization better preserves the continuous nature of our problem. Invoking results from optimal transport theory and convex algebraic geometry we reduce the semi-infinite program to a finite one and provide a practical implementation based on semidefinite programming. We show, experimentally and in theory, that the approach successfully reduces the duality gap. To showcase the scalability of our approach, we apply it to the stereo matching problem between two images.

Marcel Seelbach Benkner, Vladislav Golyanik, Christian Theobalt et al.

Matching problems on 3D shapes and images are challenging as they are frequently formulated as combinatorial quadratic assignment problems (QAPs) with permutation matrix constraints, which are NP-hard. In this work, we address such problems with emerging quantum computing technology and propose several reformulations of QAPs as unconstrained problems suitable for efficient execution on quantum hardware. We investigate several ways to inject permutation matrix constraints in a quadratic unconstrained binary optimization problem which can be mapped to quantum hardware. We focus on obtaining a sufficient spectral gap, which further increases the probability to measure optimal solutions and valid permutation matrices in a single run. We perform our experiments on the quantum computer D-Wave 2000Q (2^11 qubits, adiabatic). Despite the observed discrepancy between simulated adiabatic quantum computing and execution on real quantum hardware, our reformulation of permutation matrix constraints increases the robustness of the numerical computations over other penalty approaches in our experiments. The proposed algorithm has the potential to scale to higher dimensions on future quantum computing architectures, which opens up multiple new directions for solving matching problems in 3D computer vision and graphics.

Kanchana Vaishnavi Gandikota, Jonas Geiping, Zorah Lähner et al.

Many applications require the robustness, or ideally the invariance, of a neural network to certain transformations of input data. Most commonly, this requirement is addressed by either augmenting the training data, using adversarial training, or defining network architectures that include the desired invariance automatically. Unfortunately, the latter often relies on the ability to enlist all possible transformations, which make such approaches largely infeasible for infinite sets of transformations, such as arbitrary rotations or scaling. In this work, we propose a method for provably invariant network architectures with respect to group actions by choosing one element from a (possibly continuous) orbit based on a fixed criterion. In a nutshell, we intend to 'undo' any possible transformation before feeding the data into the actual network. We analyze properties of such approaches, extend them to equivariant networks, and demonstrate their advantages in terms of robustness as well as computational efficiency in several numerical examples. In particular, we investigate the robustness with respect to rotations of images (which can possibly hold up to discretization artifacts only) as well as the provable rotational and scaling invariance of 3D point cloud classification.

Marcel Seelbach Benkner, Zorah Lähner, Vladislav Golyanik et al.

Finding shape correspondences can be formulated as an NP-hard quadratic assignment problem (QAP) that becomes infeasible for shapes with high sampling density. A promising research direction is to tackle such quadratic optimization problems over binary variables with quantum annealing, which allows for some problems a more efficient search in the solution space. Unfortunately, enforcing the linear equality constraints in QAPs via a penalty significantly limits the success probability of such methods on currently available quantum hardware. To address this limitation, this paper proposes Q-Match, i.e., a new iterative quantum method for QAPs inspired by the alpha-expansion algorithm, which allows solving problems of an order of magnitude larger than current quantum methods. It implicitly enforces the QAP constraints by updating the current estimates in a cyclic fashion. Further, Q-Match can be applied iteratively, on a subset of well-chosen correspondences, allowing us to scale to real-world problems. Using the latest quantum annealer, the D-Wave Advantage, we evaluate the proposed method on a subset of QAPLIB as well as on isometric shape matching problems from the FAUST dataset.

Jonas Geiping, Liam Fowl, Gowthami Somepalli et al.

Data poisoning is a threat model in which a malicious actor tampers with training data to manipulate outcomes at inference time. A variety of defenses against this threat model have been proposed, but each suffers from at least one of the following flaws: they are easily overcome by adaptive attacks, they severely reduce testing performance, or they cannot generalize to diverse data poisoning threat models. Adversarial training, and its variants, are currently considered the only empirically strong defense against (inference-time) adversarial attacks. In this work, we extend the adversarial training framework to defend against (training-time) data poisoning, including targeted and backdoor attacks. Our method desensitizes networks to the effects of such attacks by creating poisons during training and injecting them into training batches. We show that this defense withstands adaptive attacks, generalizes to diverse threat models, and incurs a better performance trade-off than previous defenses such as DP-SGD or (evasion) adversarial training.

Jonas Geiping, Liam Fowl, W. Ronny Huang et al.

Data Poisoning attacks modify training data to maliciously control a model trained on such data. In this work, we focus on targeted poisoning attacks which cause a reclassification of an unmodified test image and as such breach model integrity. We consider a particularly malicious poisoning attack that is both "from scratch" and "clean label", meaning we analyze an attack that successfully works against new, randomly initialized models, and is nearly imperceptible to humans, all while perturbing only a small fraction of the training data. Previous poisoning attacks against deep neural networks in this setting have been limited in scope and success, working only in simplified settings or being prohibitively expensive for large datasets. The central mechanism of the new attack is matching the gradient direction of malicious examples. We analyze why this works, supplement with practical considerations. and show its threat to real-world practitioners, finding that it is the first poisoning method to cause targeted misclassification in modern deep networks trained from scratch on a full-sized, poisoned ImageNet dataset. Finally we demonstrate the limitations of existing defensive strategies against such an attack, concluding that data poisoning is a credible threat, even for large-scale deep learning systems.

Christina Runkel, Stefan Dorenkamp, Hartmut Bauermeister et al.

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.

Guruprasad Hegde, Avinash Nittur Ramesh, Kanchana Vaishnavi Gandikota et al.

Deep Learning systems have proven to be extremely successful for image recognition tasks for which significant amounts of training data is available, e.g., on the famous ImageNet dataset. We demonstrate that for robotics applications with cheap camera equipment, the low image quality, however,influences the classification accuracy, and freely available databases cannot be exploited in a straight forward way to train classifiers to be used on a robot. As a solution we propose to train a network on degrading the quality images in order to mimic specific low quality imaging systems. Numerical experiments demonstrate that classification networks trained by using images produced by our quality degrading network along with the high quality images outperform classification networks trained only on high quality data when used on a real robot system, while being significantly easier to use than competing zero-shot domain adaptation techniques.

Paramanand Chandramouli, Kanchana Vaishnavi Gandikota, Andreas Goerlitz et al.

Recently deep generative models have achieved impressive progress in modeling the distribution of training data. In this work, we present for the first time a generative model for 4D light field patches using variational autoencoders to capture the data distribution of light field patches. We develop a generative model conditioned on the central view of the light field and incorporate this as a prior in an energy minimization framework to address diverse light field reconstruction tasks. While pure learning-based approaches do achieve excellent results on each instance of such a problem, their applicability is limited to the specific observation model they have been trained on. On the contrary, our trained light field generative model can be incorporated as a prior into any model-based optimization approach and therefore extend to diverse reconstruction tasks including light field view synthesis, spatial-angular super resolution and reconstruction from coded projections. Our proposed method demonstrates good reconstruction, with performance approaching end-to-end trained networks, while outperforming traditional model-based approaches on both synthetic and real scenes. Furthermore, we show that our approach enables reliable light field recovery despite distortions in the input.

Jonas Geiping, Fjedor Gaede, Hartmut Bauermeister et al.

Matching and partitioning problems are fundamentals of computer vision applications with examples in multilabel segmentation, stereo estimation and optical-flow computation. These tasks can be posed as non-convex energy minimization problems and solved near-globally optimal by recent convex lifting approaches. Yet, applying these techniques comes with a significant computational effort, reducing their feasibility in practical applications. We discuss spatial discretization of continuous partitioning problems into a graph structure, generalizing discretization onto a Cartesian grid. This setup allows us to faithfully work on super-pixel graphs constructed by SLIC or Cut-Pursuit, massively decreasing the computational effort for lifted partitioning problems compared to a Cartesian grid, while optimal energy values remain similar: The global matching is still solved near-globally optimal. We discuss this methodology in detail and show examples in multi-label segmentation by minimal partitions and stereo estimation, where we demonstrate that the proposed graph discretization can reduce runtime as well as memory consumption of convex relaxations of matching problems by up to a factor of 10.

Jonas Geiping, Hartmut Bauermeister, Hannah Dröge et al.

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 in computer vision.

Micah Goldblum, Jonas Geiping, Avi Schwarzschild et al.

We empirically evaluate common assumptions about neural networks that are widely held by practitioners and theorists alike. In this work, we: (1) prove the widespread existence of suboptimal local minima in the loss landscape of neural networks, and we use our theory to find examples; (2) show that small-norm parameters are not optimal for generalization; (3) demonstrate that ResNets do not conform to wide-network theories, such as the neural tangent kernel, and that the interaction between skip connections and batch normalization plays a role; (4) find that rank does not correlate with generalization or robustness in a practical setting.

Jonas Geiping, Michael Moeller

Energy minimization methods are a classical tool in a multitude of computer vision applications. While they are interpretable and well-studied, their regularity assumptions are difficult to design by hand. Deep learning techniques on the other hand are purely data-driven, often provide excellent results, but are very difficult to constrain to predefined physical or safety-critical models. A possible combination between the two approaches is to design a parametric energy and train the free parameters in such a way that minimizers of the energy correspond to desired solution on a set of training examples. Unfortunately, such formulations typically lead to bi-level optimization problems, on which common optimization algorithms are difficult to scale to modern requirements in data processing and efficiency. In this work, we present a new strategy to optimize these bi-level problems. We investigate surrogate single-level problems that majorize the target problems and can be implemented with existing tools, leading to efficient algorithms without collapse of the energy function. This framework of strategies enables new avenues to the training of parameterized energy minimization models from large data.

Michael Moeller, Thomas Möllenhoff, Daniel Cremers

The last decade has shown a tremendous success in solving various computer vision problems with the help of deep learning techniques. Lately, many works have demonstrated that learning-based approaches with suitable network architectures even exhibit superior performance for the solution of (ill-posed) image reconstruction problems such as deblurring, super-resolution, or medical image reconstruction. The drawback of purely learning-based methods, however, is that they cannot provide provable guarantees for the trained network to follow a given data formation process during inference. In this work we propose energy dissipating networks that iteratively compute a descent direction with respect to a given cost function or energy at the currently estimated reconstruction. Therefore, an adaptive step size rule such as a line-search, along with a suitable number of iterations can guarantee the reconstruction to follow a given data formation model encoded in the energy to arbitrary precision, and hence control the model's behavior even during test time. We prove that under standard assumptions, descent using the direction predicted by the network converges (linearly) to the global minimum of the energy. We illustrate the effectiveness of the proposed approach in experiments on single image super resolution and computed tomography (CT) reconstruction, and further illustrate extensions to convex feasibility problems.

Rania Briq, Michael Moeller, Juergen Gall

Weakly supervised semantic segmentation has been a subject of increased interest due to the scarcity of fully annotated images. We introduce a new approach for solving weakly supervised semantic segmentation with deep Convolutional Neural Networks (CNNs). The method introduces a novel layer which applies simplex projection on the output of a neural network using area constraints of class objects. The proposed method is general and can be seamlessly integrated into any CNN architecture. Moreover, the projection layer allows strongly supervised models to be adapted to weakly supervised models effortlessly by substituting ground truth labels. Our experiments have shown that applying such an operation on the output of a CNN improves the accuracy of semantic segmentation in a weakly supervised setting with image-level labels.

Michael Moeller, Otmar Loffeld, Juergen Gall et al.

The idea of compressed sensing is to exploit representations in suitable (overcomplete) dictionaries that allow to recover signals far beyond the Nyquist rate provided that they admit a sparse representation in the respective dictionary. The latter gives rise to the sparse recovery problem of finding the best sparse linear approximation of given data in a given generating system. In this paper we analyze the iterative hard thresholding (IHT) algorithm as one of the most popular greedy methods for solving the sparse recovery problem, and demonstrate that systematically perturbing the IHT algorithm by adding noise to intermediate iterates yields improved results. Further improvements can be obtained by entirely rephrasing the problem as a parametric deep-learning-type of optimization problem. By introducing perturbations via dropout, we demonstrate to significantly outperform the classical IHT algorithm, obtaining $3$ to $6$ times lower average objective errors.

Peter Ochs, Tim Meinhardt, Laura Leal-Taixe et al.

The great advances of learning-based approaches in image processing and computer vision are largely based on deeply nested networks that compose linear transfer functions with suitable non-linearities. Interestingly, the most frequently used non-linearities in imaging applications (variants of the rectified linear unit) are uncommon in low dimensional approximation problems. In this paper we propose a novel non-linear transfer function, called lifting, which is motivated from a related technique in convex optimization. A lifting layer increases the dimensionality of the input, naturally yields a linear spline when combined with a fully connected layer, and therefore closes the gap between low and high dimensional approximation problems. Moreover, applying the lifting operation to the loss layer of the network allows us to handle non-convex and flat (zero-gradient) cost functions. We analyze the proposed lifting theoretically, exemplify interesting properties in synthetic experiments and demonstrate its effectiveness in deep learning approaches to image classification and denoising.