Michael Arens

Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner et al.

This paper proposes a novel method for vision-based metric cross-view geolocalization (CVGL) that matches the camera images captured from a ground-based vehicle with an aerial image to determine the vehicle's geo-pose. Since aerial images are globally available at low cost, they represent a potential compromise between two established paradigms of autonomous driving, i.e. using expensive high-definition prior maps or relying entirely on the sensor data captured at runtime. We present an end-to-end differentiable model that uses the ground and aerial images to predict a probability distribution over possible vehicle poses. We combine multiple vehicle datasets with aerial images from orthophoto providers on which we demonstrate the feasibility of our method. Since the ground truth poses are often inaccurate w.r.t. the aerial images, we implement a pseudo-label approach to produce more accurate ground truth poses and make them publicly available. While previous works require training data from the target region to achieve reasonable localization accuracy (i.e. same-area evaluation), our approach overcomes this limitation and outperforms previous results even in the strictly more challenging cross-area case. We improve the previous state-of-the-art by a large margin even without ground or aerial data from the test region, which highlights the model's potential for global-scale application. We further integrate the uncertainty-aware predictions in a tracking framework to determine the vehicle's trajectory over time resulting in a mean position error on KITTI-360 of 0.78m.

Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner et al.

This paper proposes a novel method for geo-tracking, i.e. continuous metric self-localization in outdoor environments by registering a vehicle's sensor information with aerial imagery of an unseen target region. Geo-tracking methods offer the potential to supplant noisy signals from global navigation satellite systems (GNSS) and expensive and hard to maintain prior maps that are typically used for this purpose. The proposed geo-tracking method aligns data from on-board cameras and lidar sensors with geo-registered orthophotos to continuously localize a vehicle. We train a model in a metric learning setting to extract visual features from ground and aerial images. The ground features are projected into a top-down perspective via the lidar points and are matched with the aerial features to determine the relative pose between vehicle and orthophoto. Our method is the first to utilize on-board cameras in an end-to-end differentiable model for metric self-localization on unseen orthophotos. It exhibits strong generalization, is robust to changes in the environment and requires only geo-poses as ground truth. We evaluate our approach on the KITTI-360 dataset and achieve a mean absolute position error (APE) of 0.94m. We further compare with previous approaches on the KITTI odometry dataset and achieve state-of-the-art results on the geo-tracking task.

Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner et al.

This work presents a method that is able to predict the geolocation of a street-view photo taken in the wild within a state-sized search region by matching against a database of aerial reference imagery. We partition the search region into geographical cells and train a model to map cells and corresponding photos into a joint embedding space that is used to perform retrieval at test time. The model utilizes aerial images for each cell at multiple levels-of-detail to provide sufficient information about the surrounding scene. We propose a novel layout of the search region with consistent cell resolutions that allows scaling to large geographical regions. Experiments demonstrate that the method successfully localizes 60.6% of all non-panoramic street-view photos uploaded to the crowd-sourcing platform Mapillary in the state of Massachusetts to within 50m of their ground-truth location. Source code is available at https://github.com/fferflo/statewide-visual-geolocalization.

Stephane Vujasinovic, Sebastian Bullinger, Stefan Becker et al.

While current methods for interactive Video Object Segmentation (iVOS) rely on scribble-based interactions to generate precise object masks, we propose a Click-based interactive Video Object Segmentation (CiVOS) framework to simplify the required user workload as much as possible. CiVOS builds on de-coupled modules reflecting user interaction and mask propagation. The interaction module converts click-based interactions into an object mask, which is then inferred to the remaining frames by the propagation module. Additional user interactions allow for a refinement of the object mask. The approach is extensively evaluated on the popular interactive~DAVIS dataset, but with an inevitable adaptation of scribble-based interactions with click-based counterparts. We consider several strategies for generating clicks during our evaluation to reflect various user inputs and adjust the DAVIS performance metric to perform a hardware-independent comparison. The presented CiVOS pipeline achieves competitive results, although requiring a lower user workload.

Stéphane Vujasinović, Stefan Becker, Sebastian Bullinger et al.

In this paper, we introduce a variant of video object segmentation (VOS) that bridges interactive and semi-automatic approaches, termed Lazy Video Object Segmentation (ziVOS). In contrast, to both tasks, which handle video object segmentation in an off-line manner (i.e., pre-recorded sequences), we propose through ziVOS to target online recorded sequences. Here, we strive to strike a balance between performance and robustness for long-term scenarios by soliciting user feedback's on-the-fly during the segmentation process. Hence, we aim to maximize the tracking duration of an object of interest, while requiring minimal user corrections to maintain tracking over an extended period. We propose a competitive baseline, i.e., Lazy-XMem, as a reference for future works in ziVOS. Our proposed approach uses an uncertainty estimation of the tracking state to determine whether a user interaction is necessary to refine the model's prediction. To quantitatively assess the performance of our method and the user's workload, we introduce complementary metrics alongside those already established in the field. We evaluate our approach using the recently introduced LVOS dataset, which offers numerous long-term videos. Our code is publicly available at https://github.com/Vujas-Eteph/LazyXMem.

Jens Bayer, Stefan Becker, David Münch et al.

Higher-order adversarial attacks can directly be considered the result of a cat-and-mouse game -- an elaborate action involving constant pursuit, near captures, and repeated escapes. This idiom describes the enduring circular training of adversarial attack patterns and adversarial training the best. The following work investigates the impact of higher-order adversarial attacks on object detectors by successively training attack patterns and hardening object detectors with adversarial training. The YOLOv10 object detector is chosen as a representative, and adversarial patches are used in an evasion attack manner. Our results indicate that higher-order adversarial patches are not only affecting the object detector directly trained on but rather provide a stronger generalization capacity compared to lower-order adversarial patches. Moreover, the results highlight that solely adversarial training is not sufficient to harden an object detector efficiently against this kind of adversarial attack. Code: https://github.com/JensBayer/HigherOrder

Ronny Hug, Stefan Becker, Wolfgang Hübner et al.

Probabilistic models for sequential data are the basis for a variety of applications concerned with processing timely ordered information. The predominant approach in this domain is given by recurrent neural networks, implementing either an approximate Bayesian approach (e.g. Variational Autoencoders or Generative Adversarial Networks) or a regression-based approach, i.e. variations of Mixture Density networks (MDN). In this paper, we focus on the $\mathcal{N}$-MDN variant, which parameterizes (mixtures of) probabilistic Bézier curves ($\mathcal{N}$-Curves) for modeling stochastic processes. While in favor in terms of computational cost and stability, MDNs generally fall behind approximate Bayesian approaches in terms of expressiveness. Towards this end, we present an approach for closing this gap by enabling full Bayesian inference on top of $\mathcal{N}$-MDNs. For this, we show that $\mathcal{N}$-Curves are a special case of Gaussian processes (denoted as $\mathcal{N}$-GP) and then derive corresponding mean and kernel functions for different modalities. Following this, we propose the use of the $\mathcal{N}$-MDN as a data-dependent generator for $\mathcal{N}$-GP prior distributions. We show the advantages granted by this combined model in an application context, using human trajectory prediction as an example.

Jens Bayer, Stefan Becker, David Münch et al.

Adversarial patches are still a simple yet powerful white box attack that can be used to fool object detectors by suppressing possible detections. The patches of these so-called evasion attacks are computational expensive to produce and require full access to the attacked detector. This paper addresses the problem of computational expensiveness by analyzing 375 generated patches, calculating the principal components of these and show, that linear combinations of the resulting "eigenpatches" can be used to fool object detections successfully.

Stefan Becker, Jens Bayer, Ronny Hug et al.

Data pooling offers various advantages, such as increasing the sample size, improving generalization, reducing sampling bias, and addressing data sparsity and quality, but it is not straightforward and may even be counterproductive. Assessing the effectiveness of pooling datasets in a principled manner is challenging due to the difficulty in estimating the overall information content of individual datasets. Towards this end, we propose incorporating a data source prediction module into standard object detection pipelines. The module runs with minimal overhead during inference time, providing additional information about the data source assigned to individual detections. We show the benefits of the so-called dataset affinity score by automatically selecting samples from a heterogeneous pool of vehicle datasets. The results show that object detectors can be trained on a significantly sparser set of training samples without losing detection accuracy.

Sebastian Bullinger, Florian Fervers, Christoph Bodensteiner et al.

To cope with the high requirements during the computation of semantic segmentations of earth observation imagery, current state-of-the-art pipelines divide the corresponding data into smaller images. Existing methods and benchmark datasets oftentimes rely on pixel-based tiling schemes or on geo-tiling schemes employed by web mapping applications. The selection of subimages (comprising size, location and orientation) is crucial. It affects the available context information of each pixel, defines the number of tiles during training, and influences the degree of information degradation while down- and up-sampling the tile contents to the size required by the segmentation model. We propose a new segmentation pipeline for earth observation imagery relying on a tiling scheme that creates geo-tiles based on the geo-information of the raster data. This approach exhibits several beneficial properties compared to pixel-based or common web mapping approaches. The proposed tiling scheme shows flexible customization properties regarding tile granularity, tile stride and image boundary alignment. This allows us to perform a tile specific data augmentation during training and a substitution of pixel predictions with limited context information using data of overlapping tiles during inference. The generated tiles show a consistent spatial tile extent w.r.t. heterogeneous sensors, varying recording distances and different latitudes. We demonstrate how the proposed tiling system allows to improve the results of current state-of-the-art semantic segmentation models. To foster future research we make the source code publicly available.

Jens Bayer, Stefan Becker, David Münch et al.

Adversarial patches in computer vision can be used, to fool deep neural networks and manipulate their decision-making process. One of the most prominent examples of adversarial patches are evasion attacks for object detectors. By covering parts of objects of interest, these patches suppress the detections and thus make the target object 'invisible' to the object detector. Since these patches are usually optimized on a specific network with a specific train dataset, the transferability across multiple networks and datasets is not given. This paper addresses these issues and investigates the transferability across numerous object detector architectures. Our extensive evaluation across various models on two distinct datasets indicates that patches optimized with larger models provide better network transferability than patches that are optimized with smaller models.

Valentin Wagner, Sebastian Bullinger, Michael Arens et al.

We present SatGeo-NeRF, a geometrically regularized NeRF for satellite imagery that mitigates overfitting-induced geometric artifacts observed in current state-of-the-art models using three model-agnostic regularizers. Gravity-Aligned Planarity Regularization aligns depth-inferred, approximated surface normals with the gravity axis to promote local planarity, coupling adjacent rays via a corresponding surface approximation to facilitate cross-ray gradient flow. Granularity Regularization enforces a coarse-to-fine geometry-learning scheme, and Depth-Supervised Regularization stabilizes early training for improved geometric accuracy. On the DFC2019 satellite reconstruction benchmark, SatGeo-NeRF improves the Mean Altitude Error by 13.9% and 11.7% relative to state-of-the-art baselines such as EO-NeRF and EO-GS.

Stefan Becker, Simon Weiss, Wolfgang Hübner et al.

Image degradations can occur during acquisition, processing, and transmission, altering visual appearance and affecting downstream vision tasks. They are studied in several communities, including synthetic corruption benchmarks for robustness evaluation, perceptual image quality assessment, and physically grounded analyses of imaging systems or real camera failures. Although these areas address closely related phenomena, they often use incompatible grouping schemes and backend specific severity definitions, making results difficult to compare across datasets, degradation sources, and tasks. We propose a causally grounded framework for organizing and interpreting image degradations across these settings. Instead of introducing new degradations or redefining existing benchmarks, we provide an interpretive representation and measurement layer that makes implicit assumptions explicit. Each degradation is described along two orthogonal axes: its dominant causal source in the imaging pipeline (environment, sensor/optics, ISP/renderer/codec, or transfer/system), and its resulting perceptual effect. This dual axis abstraction yields a compact taxonomy spanning algorithmic corruptions, perceptual distortions, and physically motivated imaging artifacts. To address inconsistent severity semantics without changing existing implementations, we introduce a lightweight severity measurement layer. For every degradation and each native severity level of a given backend, we quantify degradation strength using full reference image quality metrics: PSNR, SSIM, and LPIPS. This makes severity observable and comparable across sources while preserving native parameterizations. We demonstrate the framework through COCO Degradation, a taxonomy aligned benchmark for evaluating object detector robustness under diverse imaging conditions.

Valentin Wagner, Sebastian Bullinger, Christoph Bodensteiner et al.

In this work we propose a satellite specific Neural Radiance Fields (NeRF) model capable to obtain a three-dimensional semantic representation (neural semantic field) of the scene. The model derives the output from a set of multi-date satellite images with corresponding pixel-wise semantic labels. We demonstrate the robustness of our approach and its capability to improve noisy input labels. We enhance the color prediction by utilizing the semantic information to address temporal image inconsistencies caused by non-stationary categories such as vehicles. To facilitate further research in this domain, we present a dataset comprising manually generated labels for popular multi-view satellite images. Our code and dataset are available at https://github.com/wagnva/semantic-nerf-for-satellite-data.

Florian Fervers, Timo Breuer, Gregor Stachowiak et al.

Models for semantic segmentation require a large amount of hand-labeled training data which is costly and time-consuming to produce. For this purpose, we present a label fusion framework that is capable of improving semantic pixel labels of video sequences in an unsupervised manner. We make use of a 3D mesh representation of the environment and fuse the predictions of different frames into a consistent representation using semantic mesh textures. Rendering the semantic mesh using the original intrinsic and extrinsic camera parameters yields a set of improved semantic segmentation images. Due to our optimized CUDA implementation, we are able to exploit the entire $c$-dimensional probability distribution of annotations over $c$ classes in an uncertainty-aware manner. We evaluate our method on the Scannet dataset where we improve annotations produced by the state-of-the-art segmentation network ESANet from $52.05 \%$ to $58.25 \%$ pixel accuracy. We publish the source code of our framework online to foster future research in this area (\url{https://github.com/fferflo/semantic-meshes}). To the best of our knowledge, this is the first publicly available label fusion framework for semantic image segmentation based on meshes with semantic textures.

Sebastian Bullinger, Christoph Bodensteiner, Michael Arens

We propose a framework that extends Blender to exploit Structure from Motion (SfM) and Multi-View Stereo (MVS) techniques for image-based modeling tasks such as sculpting or camera and motion tracking. Applying SfM allows us to determine camera motions without manually defining feature tracks or calibrating the cameras used to capture the image data. With MVS we are able to automatically compute dense scene models, which is not feasible with the built-in tools of Blender. Currently, our framework supports several state-of-the-art SfM and MVS pipelines. The modular system design enables us to integrate further approaches without additional effort. The framework is publicly available as an open source software package.

Vanessa Buhrmester, David Muench, Dimitri Bulatov et al.

Explainable Artificial Intelligence (XAI) is increasingly rec ognized as essential for deploying machine learning systems in safety critical environments. In Automatic Target Recognition (ATR), where models operate on image, video, radar, and multisensor data, high pre dictive performance alone is insufficient. Model decisions must also be interpretable, reliable, and suitable for validation. This paper presents a structured evaluation of explainability methods in the context of safety-critical ATR systems: We identify major XAI paradigms, including saliency-based, attention-based, and surrogate ap proaches, as well as recent detection-aware extensions. Based on this, we formalize explainability as an assurance-oriented assessment problem, introduce a taxonomy, and assess these methods with respect to four key dimensions: interpretability, robustness, vulnerability to manipula tion, and suitability for validation and verification. The analysis identifies systematic limitations of current post-hoc explanation methods. In par ticular, we derive critical failure modes such as spurious explanations, instability under perturbations, and overtrust induced by visually con vincing outputs. These findings indicate that widely used XAI techniques may be insufficient for safety-critical deployment. Finally, we discuss implications for ATR systems and outline directions toward more robust, causally grounded, and physically informed explain ability methods. Our results emphasize the need to move beyond visually plausible explanations toward approaches that support reliable decision making and system-level assurance.

Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner et al.

To find the geolocation of a street-view image, cross-view geolocalization (CVGL) methods typically perform image retrieval on a database of georeferenced aerial images and determine the location from the visually most similar match. Recent approaches focus mainly on settings where street-view and aerial images are preselected to align w.r.t. translation or orientation, but struggle in challenging real-world scenarios where varying camera poses have to be matched to the same aerial image. We propose a novel trainable retrieval architecture that uses bird's eye view (BEV) maps rather than vectors as embedding representation, and explicitly addresses the many-to-one ambiguity that arises in real-world scenarios. The BEV-based retrieval is trained using the same contrastive setting and loss as classical retrieval. Our method C-BEV surpasses the state-of-the-art on the retrieval task on multiple datasets by a large margin. It is particularly effective in challenging many-to-one scenarios, e.g. increasing the top-1 recall on VIGOR's cross-area split with unknown orientation from 31.1% to 65.0%. Although the model is supervised only through a contrastive objective applied on image pairings, it additionally learns to infer the 3-DoF camera pose on the matching aerial image, and even yields a lower mean pose error than recent methods that are explicitly trained with metric groundtruth.

Jens Bayer, Stefan Becker, David Münch et al.

Despite ongoing research on the topic of adversarial examples in deep learning for computer vision, some fundamentals of the nature of these attacks remain unclear. As the manifold hypothesis posits, high-dimensional data tends to be part of a low-dimensional manifold. To verify the thesis with adversarial patches, this paper provides an analysis of a set of adversarial patches and investigates the reconstruction abilities of three different dimensionality reduction methods. Quantitatively, the performance of reconstructed patches in an attack setting is measured and the impact of sampled patches from the latent space during adversarial training is investigated. The evaluation is performed on two publicly available datasets for person detection. The results indicate that more sophisticated dimensionality reduction methods offer no advantages over a simple principal component analysis.

Stefan Becker, Simon Weiss, Wolfgang Hübner et al.

Object detectors achieve strong performance under nominal imaging conditions but can fail silently when exposed to blur, noise, compression, adverse weather, or resolution changes. In safety-critical settings, it is therefore insufficient to produce predictions without assessing whether the input remains within the detector's nominal operating regime. We refer to this capability as self-aware object detection. We introduce a degradation-aware self-awareness framework based on degradation manifolds, which explicitly structure a detector's feature space according to image degradation rather than semantic content. Our method augments a standard detection backbone with a lightweight embedding head trained via multi-layer contrastive learning. Images sharing the same degradation composition are pulled together, while differing degradation configurations are pushed apart, yielding a geometrically organized representation that captures degradation type and severity without requiring degradation labels or explicit density modeling. To anchor the learned geometry, we estimate a pristine prototype from clean training embeddings, defining a nominal operating point in representation space. Self-awareness emerges as geometric deviation from this reference, providing an intrinsic, image-level signal of degradation-induced shift that is independent of detection confidence. Extensive experiments on synthetic corruption benchmarks, cross-dataset zero-shot transfer, and natural weather-induced distribution shifts demonstrate strong pristine-degraded separability, consistent behavior across multiple detector architectures, and robust generalization under semantic shift. These results suggest that degradation-aware representation geometry provides a practical and detector-agnostic foundation.

Ronny Hug, Stefan Becker, Wolfgang Hübner et al.

An appropriate data basis grants one of the most important aspects for training and evaluating probabilistic trajectory prediction models based on neural networks. In this regard, a common shortcoming of current benchmark datasets is their limitation to sets of sample trajectories and a lack of actual ground truth distributions, which prevents the use of more expressive error metrics, such as the Wasserstein distance for model evaluation. Towards this end, this paper proposes a novel approach to synthetic dataset generation based on composite probabilistic Bézier curves, which is capable of generating ground truth data in terms of probability distributions over full trajectories. This allows the calculation of arbitrary posterior distributions. The paper showcases an exemplary trajectory prediction model evaluation using generated ground truth distribution data.

Stéphane Vujasinović, Sebastian Bullinger, Stefan Becker et al.

We present READMem (Robust Embedding Association for a Diverse Memory), a modular framework for semi-automatic video object segmentation (sVOS) methods designed to handle unconstrained videos. Contemporary sVOS works typically aggregate video frames in an ever-expanding memory, demanding high hardware resources for long-term applications. To mitigate memory requirements and prevent near object duplicates (caused by information of adjacent frames), previous methods introduce a hyper-parameter that controls the frequency of frames eligible to be stored. This parameter has to be adjusted according to concrete video properties (such as rapidity of appearance changes and video length) and does not generalize well. Instead, we integrate the embedding of a new frame into the memory only if it increases the diversity of the memory content. Furthermore, we propose a robust association of the embeddings stored in the memory with query embeddings during the update process. Our approach avoids the accumulation of redundant data, allowing us in return, to restrict the memory size and prevent extreme memory demands in long videos. We extend popular sVOS baselines with READMem, which previously showed limited performance on long videos. Our approach achieves competitive results on the Long-time Video dataset (LV1) while not hindering performance on short sequences. Our code is publicly available.

Jens Bayer, David Münch, Michael Arens

Deep neural networks, especially convolutional deep neural networks, are state-of-the-art methods to classify, segment or even generate images, movies, or sounds. However, these methods lack of a good semantic understanding of what happens internally. The question, why a COVID-19 detector has classified a stack of lung-ct images as positive, is sometimes more interesting than the overall specificity and sensitivity. Especially when human domain expert knowledge disagrees with the given output. This way, human domain experts could also be advised to reconsider their choice, regarding the information pointed out by the system. In addition, the deep learning model can be controlled, and a present dataset bias can be found. Currently, most explainable AI methods in the computer vision domain are purely used on image classification, where the images are ordinary images in the visible spectrum. As a result, there is no comparison on how the methods behave with multimodal image data, as well as most methods have not been investigated on how they behave when used for object detection. This work tries to close the gaps. Firstly, investigating three saliency map generator methods on how their maps differ across the different spectra. This is achieved via accurate and systematic training. Secondly, we examine how they behave when used for object detection. As a practical problem, we chose object detection in the infrared and visual spectrum for autonomous driving. The dataset used in this work is the Multispectral Object Detection Dataset, where each scene is available in the FIR, MIR and NIR as well as visual spectrum. The results show that there are differences between the infrared and visual activation maps. Further, an advanced training with both, the infrared and visual data not only improves the network's output, it also leads to more focused spots in the saliency maps.

Stefan Becker, Ronny Hug, Wolfgang Hübner et al.

Deep learning-based models, such as recurrent neural networks (RNNs), have been applied to various sequence learning tasks with great success. Following this, these models are increasingly replacing classic approaches in object tracking applications for motion prediction. On the one hand, these models can capture complex object dynamics with less modeling required, but on the other hand, they depend on a large amount of training data for parameter tuning. Towards this end, we present an approach for generating synthetic trajectory data of unmanned-aerial-vehicles (UAVs) in image space. Since UAVs, or rather quadrotors are dynamical systems, they can not follow arbitrary trajectories. With the prerequisite that UAV trajectories fulfill a smoothness criterion corresponding to a minimal change of higher-order motion, methods for planning aggressive quadrotors flights can be utilized to generate optimal trajectories through a sequence of 3D waypoints. By projecting these maneuver trajectories, which are suitable for controlling quadrotors, to image space, a versatile trajectory data set is realized. To demonstrate the applicability of the synthetic trajectory data, we show that an RNN-based prediction model solely trained on the generated data can outperform classic reference models on a real-world UAV tracking dataset. The evaluation is done on the publicly available ANTI-UAV dataset.

Stefan Becker, Ronny Hug, Wolfgang Hübner et al.

In applications such as object tracking, time-series data inevitably carry missing observations. Following the success of deep learning-based models for various sequence learning tasks, these models increasingly replace classic approaches in object tracking applications for inferring the objects' motion states. While traditional tracking approaches can deal with missing observations, most of their deep counterparts are, by default, not suited for this. Towards this end, this paper introduces a transformer-based approach for handling missing observations in variable input length trajectory data. The model is formed indirectly by successively increasing the complexity of the demanded inference tasks. Starting from reproducing noise-free trajectories, the model then learns to infer trajectories from noisy inputs. By providing missing tokens, binary-encoded missing events, the model learns to in-attend to missing data and infers a complete trajectory conditioned on the remaining inputs. In the case of a sequence of successive missing events, the model then acts as a pure prediction model. The abilities of the approach are demonstrated on synthetic data and real-world data reflecting prototypical object tracking scenarios.

Björn Borgmann, Volker Schatz, Marcus Hammer et al.

We present the current state of development of the sensor-equipped car MODISSA, with which Fraunhofer IOSB realizes a configurable experimental platform for hardware evaluation and software development in the context of mobile mapping and vehicle-related safety and protection. MODISSA is based on a van that has successively been equipped with a variety of optical sensors over the past few years, and contains hardware for complete raw data acquisition, georeferencing, real-time data analysis, and immediate visualization on in-car displays. We demonstrate the capabilities of MODISSA by giving a deeper insight into experiments with its specific configuration in the scope of three different applications. Other research groups can benefit from these experiences when setting up their own mobile sensor system, especially regarding the selection of hardware and software, the knowledge of possible sources of error, and the handling of the acquired sensor data.

Stefan Becker, Ronny Hug, Wolfgang Hübner et al.

In tasks such as tracking, time-series data inevitably carry missing observations. While traditional tracking approaches can handle missing observations, recurrent neural networks (RNNs) are designed to receive input data in every step. Furthermore, current solutions for RNNs, like omitting the missing data or data imputation, are not sufficient to account for the resulting increased uncertainty. Towards this end, this paper introduces an RNN-based approach that provides a full temporal filtering cycle for motion state estimation. The Kalman filter inspired approach, enables to deal with missing observations and outliers. For providing a full temporal filtering cycle, a basic RNN is extended to take observations and the associated belief about its accuracy into account for updating the current state. An RNN prediction model, which generates a parametrized distribution to capture the predicted states, is combined with an RNN update model, which relies on the prediction model output and the current observation. By providing the model with masking information, binary-encoded missing events, the model can overcome limitations of standard techniques for dealing with missing input values. The model abilities are demonstrated on synthetic data reflecting prototypical pedestrian tracking scenarios.

Sebastian Bullinger, Christoph Bodensteiner, Michael Arens

The reconstruction of accurate three-dimensional environment models is one of the most fundamental goals in the field of photogrammetry. Since satellite images provide suitable properties for obtaining large-scale environment reconstructions, there exist a variety of Stereo Matching based methods to reconstruct point clouds for satellite image pairs. Recently, the first Structure from Motion (SfM) based approach has been proposed, which allows to reconstruct point clouds from multiple satellite images. In this work, we propose an extension of this SfM based pipeline that allows us to reconstruct not only point clouds but watertight meshes including texture information. We provide a detailed description of several steps that are mandatory to exploit state-of-the-art mesh reconstruction algorithms in the context of satellite imagery. This includes a decomposition of finite projective camera calibration matrices, a skew correction of corresponding depth maps and input images as well as the recovery of real-world depth maps from reparameterized depth values. The paper presents an extensive quantitative evaluation on multi-date satellite images demonstrating that the proposed pipeline combined with current meshing algorithms outperforms state-of-the-art point cloud reconstruction algorithms in terms of completeness and median error. We make the source code of our pipeline publicly available.

Stéphane Vujasinović, Stefan Becker, Timo Breuer et al.

Single visual object tracking from an unmanned aerial vehicle (UAV) poses fundamental challenges such as object occlusion, small-scale objects, background clutter, and abrupt camera motion. To tackle these difficulties, we propose to integrate the 3D structure of the observed scene into a detection-by-tracking algorithm. We introduce a pipeline that combines a model-free visual object tracker, a sparse 3D reconstruction, and a state estimator. The 3D reconstruction of the scene is computed with an image-based Structure-from-Motion (SfM) component that enables us to leverage a state estimator in the corresponding 3D scene during tracking. By representing the position of the target in 3D space rather than in image space, we stabilize the tracking during ego-motion and improve the handling of occlusions, background clutter, and small-scale objects. We evaluated our approach on prototypical image sequences, captured from a UAV with low-altitude oblique views. For this purpose, we adapted an existing dataset for visual object tracking and reconstructed the observed scene in 3D. The experimental results demonstrate that the proposed approach outperforms methods using plain visual cues as well as approaches leveraging image-space-based state estimations. We believe that our approach can be beneficial for traffic monitoring, video surveillance, and navigation.

Ronny Hug, Stefan Becker, Wolfgang Hübner et al.

Methods to quantify the complexity of trajectory datasets are still a missing piece in benchmarking human trajectory prediction models. In order to gain a better understanding of the complexity of trajectory prediction tasks and following the intuition, that more complex datasets contain more information, an approach for quantifying the amount of information contained in a dataset from a prototype-based dataset representation is proposed. The dataset representation is obtained by first employing a non-trivial spatial sequence alignment, which enables a subsequent learning vector quantization (LVQ) stage. A large-scale complexity analysis is conducted on several human trajectory prediction benchmarking datasets, followed by a brief discussion on indications for human trajectory prediction and benchmarking.

Ronny Hug, Stefan Becker, Wolfgang Hübner et al.

The analysis and quantification of sequence complexity is an open problem frequently encountered when defining trajectory prediction benchmarks. In order to enable a more informative assembly of a data basis, an approach for determining a dataset representation in terms of a small set of distinguishable prototypical sub-sequences is proposed. The approach employs a sequence alignment followed by a learning vector quantization (LVQ) stage. A first proof of concept on synthetically generated and real-world datasets shows the viability of the approach.

Jens Bayer, David Münch, Michael Arens

Living in a complex world like ours makes it unacceptable that a practical implementation of a machine learning system assumes a closed world. Therefore, it is necessary for such a learning-based system in a real world environment, to be aware of its own capabilities and limits and to be able to distinguish between confident and unconfident results of the inference, especially if the sample cannot be explained by the underlying distribution. This knowledge is particularly essential in safety-critical environments and tasks e.g. self-driving cars or medical applications. Towards this end, we transfer image-based Out-of-Distribution (OoD)-methods to graph-based data and show the applicability in action recognition. The contribution of this work is (i) the examination of the portability of recent image-based OoD-detectors for graph-based input data, (ii) a Metric Learning-based approach to detect OoD-samples, and (iii) the introduction of a novel semi-synthetic action recognition dataset. The evaluation shows that image-based OoD-methods can be applied to graph-based data. Additionally, there is a gap between the performance on intraclass and intradataset results. First methods as the examined baseline or ODIN provide reasonable results. More sophisticated network architectures - in contrast to their image-based application - were surpassed in the intradataset comparison and even lead to less classification accuracy.

Vanessa Buhrmester, David Münch, Michael Arens

Deep Learning is a state-of-the-art technique to make inference on extensive or complex data. As a black box model due to their multilayer nonlinear structure, Deep Neural Networks are often criticized to be non-transparent and their predictions not traceable by humans. Furthermore, the models learn from artificial datasets, often with bias or contaminated discriminating content. Through their increased distribution, decision-making algorithms can contribute promoting prejudge and unfairness which is not easy to notice due to lack of transparency. Hence, scientists developed several so-called explanators or explainers which try to point out the connection between input and output to represent in a simplified way the inner structure of machine learning black boxes. In this survey we differ the mechanisms and properties of explaining systems for Deep Neural Networks for Computer Vision tasks. We give a comprehensive overview about taxonomy of related studies and compare several survey papers that deal with explainability in general. We work out the drawbacks and gaps and summarize further research ideas.

Ronny Hug, Wolfgang Hübner, Michael Arens

Representations of sequential data are commonly based on the assumption that observed sequences are realizations of an unknown underlying stochastic process, where the learning problem includes determination of the model parameters. In this context the model must be able to capture the multi-modal nature of the data, without blurring between modes. This property is essential for applications like trajectory prediction or human motion modeling. Towards this end, a neural network model for continuous-time stochastic processes usable for sequence prediction is proposed. The model is based on Mixture Density Networks using Bézier curves with Gaussian random variables as control points (abbrev.: $\mathcal{N}$-Curves). Key advantages of the model include the ability of generating smooth multi-mode predictions in a single inference step which reduces the need for Monte Carlo simulation, as required in many multi-step prediction models, based on state-of-the-art neural networks. Essential properties of the proposed approach are illustrated by several toy examples and the task of multi-step sequence prediction. Further, the model performance is evaluated on two real world use-cases, i.e. human trajectory prediction and human motion modeling, outperforming different state-of-the-art models.

Patrick Schlosser, David Münch, Michael Arens

In this paper, several variants of two-stream architectures for temporal action proposal generation in long, untrimmed videos are presented. Inspired by the recent advances in the field of human action recognition utilizing 3D convolutions in combination with two-stream networks and based on the Single-Stream Temporal Action Proposals (SST) architecture, four different two-stream architectures utilizing sequences of images on one stream and sequences of images of optical flow on the other stream are subsequently investigated. The four architectures fuse the two separate streams at different depths in the model; for each of them, a broad range of parameters is investigated systematically as well as an optimal parametrization is empirically determined. The experiments on the THUMOS'14 dataset show that all four two-stream architectures are able to outperform the original single-stream SST and achieve state of the art results. Additional experiments revealed that the improvements are not restricted to a single method of calculating optical flow by exchanging the formerly used method of Brox with FlowNet2 and still achieving improvements.

Stefan Becker, Ronny Hug, Wolfgang Hübner et al.

The problem of varying dynamics of tracked objects, such as pedestrians, is traditionally tackled with approaches like the Interacting Multiple Model (IMM) filter using a Bayesian formulation. By following the current trend towards using deep neural networks, in this paper an RNN-based IMM filter surrogate is presented. Similar to an IMM filter solution, the presented RNN-based model assigns a probability value to a performed dynamic and, based on them, puts out a multi-modal distribution over future pedestrian trajectories. The evaluation is done on synthetic data, reflecting prototypical pedestrian maneuvers.

Nikolas Hesse, Sergi Pujades, Michael J. Black et al.

Statistical models of the human body surface are generally learned from thousands of high-quality 3D scans in predefined poses to cover the wide variety of human body shapes and articulations. Acquisition of such data requires expensive equipment, calibration procedures, and is limited to cooperative subjects who can understand and follow instructions, such as adults. We present a method for learning a statistical 3D Skinned Multi-Infant Linear body model (SMIL) from incomplete, low-quality RGB-D sequences of freely moving infants. Quantitative experiments show that SMIL faithfully represents the RGB-D data and properly factorizes the shape and pose of the infants. To demonstrate the applicability of SMIL, we fit the model to RGB-D sequences of freely moving infants and show, with a case study, that our method captures enough motion detail for General Movements Assessment (GMA), a method used in clinical practice for early detection of neurodevelopmental disorders in infants. SMIL provides a new tool for analyzing infant shape and movement and is a step towards an automated system for GMA.

Sebastian Bullinger, Christoph Bodensteiner, Michael Arens et al.

We present a method to reconstruct the three-dimensional trajectory of a moving instance of a known object category using stereo video data. We track the two-dimensional shape of objects on pixel level exploiting instance-aware semantic segmentation techniques and optical flow cues. We apply Structure from Motion (SfM) techniques to object and background images to determine for each frame initial camera poses relative to object instances and background structures. We refine the initial SfM results by integrating stereo camera constraints exploiting factor graphs. We compute the object trajectory by combining object and background camera pose information. In contrast to stereo matching methods, our approach leverages temporal adjacent views for object point triangulation. As opposed to monocular trajectory reconstruction approaches, our method shows no degenerated cases. We evaluate our approach using publicly available video data of vehicles in urban scenes.

Stefan Becker, Ronny Hug, Wolfgang Hübner et al.

In recent years, there is a shift from modeling the tracking problem based on Bayesian formulation towards using deep neural networks. Towards this end, in this paper the effectiveness of various deep neural networks for predicting future pedestrian paths are evaluated. The analyzed deep networks solely rely, like in the traditional approaches, on observed tracklets without human-human interaction information. The evaluation is done on the publicly available TrajNet benchmark dataset, which builds up a repository of considerable and popular datasets for trajectory-based activity forecasting. We show that a Recurrent-Encoder with a Dense layer stacked on top, referred to as RED-predictor, is able to achieve sophisticated results compared to elaborated models in such scenarios. Further, we investigate failure cases and give explanations for observed phenomena and give some recommendations for overcoming demonstrated shortcomings.

Ronny Hug, Stefan Becker, Wolfgang Hübner et al.

Recurrent neural networks are able to learn complex long-term relationships from sequential data and output a pdf over the state space. Therefore, recurrent models are a natural choice to address path prediction tasks, where a trained model is used to generate future expectations from past observations. When applied to security applications, like predicting the path of pedestrians for risk assessment, a point-wise greedy (ML) evaluation of the output pdf is not feasible, since the environment often allows multiple choices. Therefore, a robust risk assessment has to take all options into account, even if they are overall not very likely. Towards this end, a combination of particle filter sampling strategies and a LSTM-MDL model is proposed to address a multi-modal path prediction task. The capabilities and viability of the proposed approach are evaluated on several synthetic test conditions, yielding the counter-intuitive result that the simplest approach performs best. Further, the feasibility of the proposed approach is illustrated on several real world scenes.

Sebastian Bullinger, Christoph Bodensteiner, Michael Arens et al.

We present a method to reconstruct the three-dimensional trajectory of a moving instance of a known object category in monocular video data. We track the two-dimensional shape of objects on pixel level exploiting instance-aware semantic segmentation techniques and optical flow cues. We apply Structure from Motion techniques to object and background images to determine for each frame camera poses relative to object instances and background structures. By combining object and background camera pose information, we restrict the object trajectory to a one-parameter family of possible solutions. We compute a ground representation by fusing background structures and corresponding semantic segmentations. This allows us to determine an object trajectory consistent to image observations and reconstructed environment model. Our method is robust to occlusion and handles temporarily stationary objects. We show qualitative results using drone imagery. Due to the lack of suitable benchmark datasets we present a new dataset to evaluate the quality of reconstructed three-dimensional object trajectories. The video sequences contain vehicles in urban areas and are rendered using the path-tracing render engine Cycles to achieve realistic results. We perform a quantitative evaluation of the presented approach using this dataset. Our algorithm achieves an average reconstruction-to-ground-truth distance of 0.31 meter.

Sebastian Bullinger, Christoph Bodensteiner, Michael Arens

We present a method to perform online Multiple Object Tracking (MOT) of known object categories in monocular video data. Current Tracking-by-Detection MOT approaches build on top of 2D bounding box detections. In contrast, we exploit state-of-the-art instance aware semantic segmentation techniques to compute 2D shape representations of target objects in each frame. We predict position and shape of segmented instances in subsequent frames by exploiting optical flow cues. We define an affinity matrix between instances of subsequent frames which reflects locality and visual similarity. The instance association is solved by applying the Hungarian method. We evaluate different configurations of our algorithm using the MOT 2D 2015 train dataset. The evaluation shows that our tracking approach is able to track objects with high relative motions. In addition, we provide results of our approach on the MOT 2D 2015 test set for comparison with previous works. We achieve a MOTA score of 32.1.