Andreas Zell

Yuzhi Lai, Shenghai Yuan, Peizheng Li et al.

ffective Human-Robot Interaction (HRI) is crucial for enhancing accessibility and usability in real-world robotics applications. However, existing solutions often rely on gesture- only or language-only commands, making interaction inefficient and ambiguous, particularly for users with physical impairments. In this paper, we introduce FAM-HRI, an efficient multimodal framework for HRI that integrates language and gaze inputs via foundation models. By leveraging lightweight Meta ARIA glasses, our system captures real-time multimodal signals and utilizes large language models (LLMs) to fuse user intention with scene context, enabling intuitive and precise robot manipulation. Our method accurately determines the gaze fixation time interval, reducing noise caused by the gaze dynamic nature. Experimental evaluations demonstrate that FAM-HRI achieves a high success rate in task execution while maintaining a low interaction time, providing a practical solution for individuals with limited physical mobility or motor impairments. To support the community, we have released our system design, algorithms, and solutions at https://github.com/laiyuzhi/FAM-HRI.

Faris Janjoš, Marcel Hallgarten, Anthony Knittel et al.

The CVAE is one of the most widely-used models in trajectory prediction for AD. It captures the interplay between a driving context and its ground-truth future into a probabilistic latent space and uses it to produce predictions. In this paper, we challenge key components of the CVAE. We leverage recent advances in the space of the VAE, the foundation of the CVAE, which show that a simple change in the sampling procedure can greatly benefit performance. We find that unscented sampling, which draws samples from any learned distribution in a deterministic manner, can naturally be better suited to trajectory prediction than potentially dangerous random sampling. We go further and offer additional improvements including a more structured Gaussian mixture latent space, as well as a novel, potentially more expressive way to do inference with CVAEs. We show wide applicability of our models by evaluating them on the INTERACTION prediction dataset, outperforming the state of the art, as well as at the task of image modeling on the CelebA dataset, outperforming the baseline vanilla CVAE. Code is available at https://github.com/boschresearch/cuae-prediction.

Thomas Gossard, Jonas Tebbe, Andreas Ziegler et al.

Spin plays a considerable role in table tennis, making a shot's trajectory harder to read and predict. However, the spin is challenging to measure because of the ball's high velocity and the magnitude of the spin values. Existing methods either require extremely high framerate cameras or are unreliable because they use the ball's logo, which may not always be visible. Because of this, many table tennis-playing robots ignore the spin, which severely limits their capabilities. This paper proposes an easily implementable and reliable spin estimation method. We developed a dotted-ball orientation estimation (DOE) method, that can then be used to estimate the spin. The dots are first localized on the image using a CNN and then identified using geometric hashing. The spin is finally regressed from the estimated orientations. Using our algorithm, the ball's orientation can be estimated with a mean error of 2.4° and the spin estimation has an relative error lower than 1%. Spins up to 175 rps are measurable with a camera of 350 fps in real time. Using our method, we generated a dataset of table tennis ball trajectories with position and spin, available on our project page.

Benjamin Kiefer, Andreas Zell

Unmanned aerial vehicles assist in maritime search and rescue missions by flying over large search areas to autonomously search for objects or people. Reliably detecting objects of interest requires fast models to employ on embedded hardware. Moreover, with increasing distance to the ground station only part of the video data can be transmitted. In this work, we consider the problem of finding meaningful region of interest proposals in a video stream on an embedded GPU. Current object or anomaly detectors are not suitable due to their slow speed, especially on limited hardware and for large image resolutions. Lastly, objects of interest, such as pieces of wreckage, are often not known a priori. Therefore, we propose an end-to-end future frame prediction model running in real-time on embedded GPUs to generate region proposals. We analyze its performance on large-scale maritime data sets and demonstrate its benefits over traditional and modern methods.

Benjamin Kiefer, Yitong Quan, Andreas Zell

This paper introduces a novel approach to video object detection detection and tracking on Unmanned Aerial Vehicles (UAVs). By incorporating metadata, the proposed approach creates a memory map of object locations in actual world coordinates, providing a more robust and interpretable representation of object locations in both, image space and the real world. We use this representation to boost confidences, resulting in improved performance for several temporal computer vision tasks, such as video object detection, short and long-term single and multi-object tracking, and video anomaly detection. These findings confirm the benefits of metadata in enhancing the capabilities of UAVs in the field of temporal computer vision and pave the way for further advancements in this area.

Leon Amadeus Varga, Sebastian Koch, Andreas Zell

An object detection pipeline comprises a camera that captures the scene and an object detector that processes these images. The quality of the images directly affects the performance of the object detector. Many works nowadays focus either on improving the image quality or improving the object detection models independently, but neglect the importance of joint optimization of the two subsystems. The goal of this paper is to tune the detection throughput and accuracy of existing object detectors in the remote sensing scenario by focusing on optimizing the input images tailored to the object detector. To achieve this, we empirically analyze the influence of two selected camera calibration parameters (camera distortion correction and gamma correction) and five image parameters (quantization, compression, resolution, color model, additional channels) for these applications. For our experiments, we utilize three UAV data sets from different domains and a mixture of large and small state-of-the-art object detector models to provide an extensive evaluation of the influence of the pipeline parameters. Finally, we realize an object detection pipeline prototype on an embedded platform for an UAV and give a best practice recommendation for building object detection pipelines based on our findings. We show that not all parameters have an equal impact on detection accuracy and data throughput, and that by using a suitable compromise between parameters we are able to achieve higher detection accuracy for lightweight object detection models, while keeping the same data throughput.

Andreas Ziegler, Daniel Teigland, Jonas Tebbe et al.

Event cameras are becoming increasingly popular in robotics and computer vision due to their beneficial properties, e.g., high temporal resolution, high bandwidth, almost no motion blur, and low power consumption. However, these cameras remain expensive and scarce in the market, making them inaccessible to the majority. Using event simulators minimizes the need for real event cameras to develop novel algorithms. However, due to the computational complexity of the simulation, the event streams of existing simulators cannot be generated in real-time but rather have to be pre-calculated from existing video sequences or pre-rendered and then simulated from a virtual 3D scene. Although these offline generated event streams can be used as training data for learning tasks, all response time dependent applications cannot benefit from these simulators yet, as they still require an actual event camera. This work proposes simulation methods that improve the performance of event simulation by two orders of magnitude (making them real-time capable) while remaining competitive in the quality assessment.

Benjamin Kiefer, Timon Höfer, Andreas Zell

Yaw estimation of boats from the viewpoint of unmanned aerial vehicles (UAVs) and unmanned surface vehicles (USVs) or boats is a crucial task in various applications such as 3D scene rendering, trajectory prediction, and navigation. However, the lack of literature on yaw estimation of objects from the viewpoint of UAVs has motivated us to address this domain. In this paper, we propose a method based on HyperPosePDF for predicting the orientation of boats in the 6D space. For that, we use existing datasets, such as PASCAL3D+ and our own datasets, SeaDronesSee-3D and BOArienT, which we annotated manually. We extend HyperPosePDF to work in video-based scenarios, such that it yields robust orientation predictions across time. Naively applying HyperPosePDF on video data yields single-point predictions, resulting in far-off predictions and often incorrect symmetric orientations due to unseen or visually different data. To alleviate this issue, we propose aggregating the probability distributions of pose predictions, resulting in significantly improved performance, as shown in our experimental evaluation. Our proposed method could significantly benefit downstream tasks in marine robotics.

Daniel Weber, Wolfgang Fuhl, Andreas Zell et al.

In human-robot collaboration, one challenging task is to teach a robot new yet unknown objects enabling it to interact with them. Thereby, gaze can contain valuable information. We investigate if it is possible to detect objects (object or no object) merely from gaze data and determine their bounding box parameters. For this purpose, we explore different sizes of temporal windows, which serve as a basis for the computation of heatmaps, i.e., the spatial distribution of the gaze data. Additionally, we analyze different grid sizes of these heatmaps, and demonstrate the functionality in a proof of concept using different machine learning techniques. Our method is characterized by its speed and resource efficiency compared to conventional object detectors. In order to generate the required data, we conducted a study with five subjects who could move freely and thus, turn towards arbitrary objects. This way, we chose a scenario for our data collection that is as realistic as possible. Since the subjects move while facing objects, the heatmaps also contain gaze data trajectories, complicating the detection and parameter regression. We make our data set publicly available to the research community for download.

Leon Amadeus Varga, Martin Messmer, Nuri Benbarka et al.

Deep Learning could drastically boost the classification accuracy for Hyperspectral Imaging (HSI). Still, the training on the mostly small hyperspectral data sets is not trivial. Two key challenges are the large channel dimension of the recordings and the incompatibility between cameras of different manufacturers. By introducing a suitable model bias and continuously defining the channel dimension, we propose a 2D convolution optimized for these challenges of Hyperspectral Imaging. We evaluate the method based on two different hyperspectral applications (inline inspection and remote sensing). Besides the shown superiority of the model, the modification adds additional explanatory power. In addition, the model learns the necessary camera filters in a data-driven manner. Based on these camera filters, an optimal camera can be designed.

Andreas Ziegler, Thomas Gossard, Karl Vetter et al.

In recent years, robotic table tennis has become a popular research challenge for perception and robot control. Here, we present an improved table tennis robot system with high accuracy vision detection and fast robot reaction. Based on previous work, our system contains a KUKA robot arm with 6 DOF, with four frame-based cameras and two additional event-based cameras. We developed a novel calibration approach to calibrate this multimodal perception system. For table tennis, spin estimation is crucial. Therefore, we introduced a novel, and more accurate spin estimation approach. Finally, we show how combining the output of an event-based camera and a Spiking Neural Network (SNN) can be used for accurate ball detection.

Thomas Gossard, Andreas Ziegler, Levin Kolmar et al.

Accurate calibration is crucial for using multiple cameras to triangulate the position of objects precisely. However, it is also a time-consuming process that needs to be repeated for every displacement of the cameras. The standard approach is to use a printed pattern with known geometry to estimate the intrinsic and extrinsic parameters of the cameras. The same idea can be applied to event-based cameras, though it requires extra work. By using frame reconstruction from events, a printed pattern can be detected. A blinking pattern can also be displayed on a screen. Then, the pattern can be directly detected from the events. Such calibration methods can provide accurate intrinsic calibration for both frame- and event-based cameras. However, using 2D patterns has several limitations for multi-camera extrinsic calibration, with cameras possessing highly different points of view and a wide baseline. The 2D pattern can only be detected from one direction and needs to be of significant size to compensate for its distance to the camera. This makes the extrinsic calibration time-consuming and cumbersome. To overcome these limitations, we propose eWand, a new method that uses blinking LEDs inside opaque spheres instead of a printed or displayed pattern. Our method provides a faster, easier-to-use extrinsic calibration approach that maintains high accuracy for both event- and frame-based cameras.

Peizheng Li, Shuxiao Ding, You Zhou et al.

Open-world 3D semantic occupancy prediction aims to generate a voxelized 3D representation from sensor inputs while recognizing both known and unknown objects. Transferring open-vocabulary knowledge from vision-language models (VLMs) offers a promising direction but remains challenging. However, methods based on VLM-derived 2D pseudo-labels with traditional supervision are limited by a predefined label space and lack general prediction capabilities. Direct alignment with pretrained image embeddings, on the other hand, often fails to achieve reliable performance because of inconsistent image and text representations in VLMs. To address these challenges, we propose AGO, a novel 3D occupancy prediction framework with adaptive grounding to handle diverse open-world scenarios. AGO first encodes surrounding images and class prompts into 3D and text embeddings, respectively, leveraging similarity-based grounding training with 3D pseudo-labels. Additionally, a modality adapter maps 3D embeddings into a space aligned with VLM-derived image embeddings, reducing modality gaps. Experiments on Occ3D-nuScenes show that AGO improves unknown object prediction in zero-shot and few-shot transfer while achieving state-of-the-art closed-world self-supervised performance, surpassing prior methods by 4.09 mIoU. Code is available at: https://github.com/EdwardLeeLPZ/AGO.

Hannah Frank, Leon Amadeus Varga, Andreas Zell

Hyperspectral Imaging (HSI) serves as a non-destructive spatial spectroscopy technique with a multitude of potential applications. However, a recurring challenge lies in the limited size of the target datasets, impeding exhaustive architecture search. Consequently, when venturing into novel applications, reliance on established methodologies becomes commonplace, in the hope that they exhibit favorable generalization characteristics. Regrettably, this optimism is often unfounded due to the fine-tuned nature of models tailored to specific HSI contexts. To address this predicament, this study introduces an innovative benchmark dataset encompassing three markedly distinct HSI applications: food inspection, remote sensing, and recycling. This comprehensive dataset affords a finer assessment of hyperspectral model capabilities. Moreover, this benchmark facilitates an incisive examination of prevailing state-of-the-art techniques, consequently fostering the evolution of superior methodologies. Furthermore, the enhanced diversity inherent in the benchmark dataset underpins the establishment of a pretraining pipeline for HSI. This pretraining regimen serves to enhance the stability of training processes for larger models. Additionally, a procedural framework is delineated, offering insights into the handling of applications afflicted by limited target dataset sizes.

Yuzhi Lai, Shenghai Yuan, Peizheng Li et al.

We present SEER-VAR, a novel framework for egocentric vehicle-based augmented reality (AR) that unifies semantic decomposition, Context-Aware SLAM Branches (CASB), and LLM-driven recommendation. Unlike existing systems that assume static or single-view settings, SEER-VAR dynamically separates cabin and road scenes via depth-guided vision-language grounding. Two SLAM branches track egocentric motion in each context, while a GPT-based module generates context-aware overlays such as dashboard cues and hazard alerts. To support evaluation, we introduce EgoSLAM-Drive, a real-world dataset featuring synchronized egocentric views, 6DoF ground-truth poses, and AR annotations across diverse driving scenarios. Experiments demonstrate that SEER-VAR achieves robust spatial alignment and perceptually coherent AR rendering across varied environments. As one of the first to explore LLM-based AR recommendation in egocentric driving, we address the lack of comparable systems through structured prompting and detailed user studies. Results show that SEER-VAR enhances perceived scene understanding, overlay relevance, and driver ease, providing an effective foundation for future research in this direction. Code and dataset will be made open source.

Benjamin Kiefer, David Ott, Andreas Zell

Acquiring data to train deep learning-based object detectors on Unmanned Aerial Vehicles (UAVs) is expensive, time-consuming and may even be prohibited by law in specific environments. On the other hand, synthetic data is fast and cheap to access. In this work, we explore the potential use of synthetic data in object detection from UAVs across various application environments. For that, we extend the open-source framework DeepGTAV to work for UAV scenarios. We capture various large-scale high-resolution synthetic data sets in several domains to demonstrate their use in real-world object detection from UAVs by analyzing multiple training strategies across several models. Furthermore, we analyze several different data generation and sampling parameters to provide actionable engineering advice for further scientific research. The DeepGTAV framework is available at https://git.io/Jyf5j.

Faranak Shamsafar, Andreas Zell

Stereo vision is an effective technique for depth estimation with broad applicability in autonomous urban and highway driving. While various deep learning-based approaches have been developed for stereo, the input data from a binocular setup with a fixed baseline are limited. Addressing such a problem, we present an end-to-end network for processing the data from a trinocular setup, which is a combination of a narrow and a wide stereo pair. In this design, two pairs of binocular data with a common reference image are treated with shared weights of the network and a mid-level fusion. We also propose a Guided Addition method for merging the 4D data of the two baselines. Additionally, an iterative sequential self-supervised and supervised learning on real and synthetic datasets is presented, making the training of the trinocular system practical with no need to ground-truth data of the real dataset. Experimental results demonstrate that the trinocular disparity network surpasses the scenario where individual pairs are fed into a similar architecture. Code and dataset: https://github.com/cogsys-tuebingen/tristereonet.

Faranak Shamsafar, Samuel Woerz, Rafia Rahim et al.

Recent methods in stereo matching have continuously improved the accuracy using deep models. This gain, however, is attained with a high increase in computation cost, such that the network may not fit even on a moderate GPU. This issue raises problems when the model needs to be deployed on resource-limited devices. For this, we propose two light models for stereo vision with reduced complexity and without sacrificing accuracy. Depending on the dimension of cost volume, we design a 2D and a 3D model with encoder-decoders built from 2D and 3D convolutions, respectively. To this end, we leverage 2D MobileNet blocks and extend them to 3D for stereo vision application. Besides, a new cost volume is proposed to boost the accuracy of the 2D model, making it performing close to 3D networks. Experiments show that the proposed 2D/3D networks effectively reduce the computational expense (27%/95% and 72%/38% fewer parameters/operations in 2D and 3D models, respectively) while upholding the accuracy. Our code is available at https://github.com/cogsys-tuebingen/mobilestereonet.

Nuri Benbarka, Jona Schröder, Andreas Zell

Multi-object tracking is a critical component in autonomous navigation, as it provides valuable information for decision-making. Many researchers tackled the 3D multi-object tracking task by filtering out the frame-by-frame 3D detections; however, their focus was mainly on finding useful features or proper matching metrics. Our work focuses on a neglected part of the tracking system: score refinement and tracklet termination. We show that manipulating the scores depending on time consistency while terminating the tracklets depending on the tracklet score improves tracking results. We do this by increasing the matched tracklets' score with score update functions and decreasing the unmatched tracklets' score. Compared to count-based methods, our method consistently produces better AMOTA and MOTA scores when utilizing various detectors and filtering algorithms on different datasets. The improvements in AMOTA score went up to 1.83 and 2.96 in MOTA. We also used our method as a late-fusion ensembling method, and it performed better than voting-based ensemble methods by a solid margin. It achieved an AMOTA score of 67.6 on nuScenes test evaluation, which is comparable to other state-of-the-art trackers. Code is publicly available at: \url{https://github.com/cogsys-tuebingen/CBMOT}.

Hamd ul Moqeet Riaz, Nuri Benbarka, Andreas Zell

We present FourierNet, a single shot, anchor-free, fully convolutional instance segmentation method that predicts a shape vector. Consequently, this shape vector is converted into the masks' contour points using a fast numerical transform. Compared to previous methods, we introduce a new training technique, where we utilize a differentiable shape decoder, which manages the automatic weight balancing of the shape vector's coefficients. We used the Fourier series as a shape encoder because of its coefficient interpretability and fast implementation. FourierNet shows promising results compared to polygon representation methods, achieving 30.6 mAP on the MS COCO 2017 benchmark. At lower image resolutions, it runs at 26.6 FPS with 24.3 mAP. It reaches 23.3 mAP using just eight parameters to represent the mask (note that at least four parameters are needed for bounding box prediction only). Qualitative analysis shows that suppressing a reasonable proportion of higher frequencies of Fourier series, still generates meaningful masks. These results validate our understanding that lower frequency components hold higher information for the segmentation task, and therefore, we can achieve a compressed representation. Code is available at: github.com/cogsys-tuebingen/FourierNet.

Kevin Alexander Laube, Andreas Zell

While recent NAS algorithms are thousands of times faster than the pioneering works, it is often overlooked that they use fewer candidate operations, resulting in a significantly smaller search space. We present PR-DARTS, a NAS algorithm that discovers strong network configurations in a much larger search space and a single day. A small candidate operation pool is used, from which candidates are progressively pruned and replaced with better performing ones. Experiments on CIFAR-10 and CIFAR-100 achieve 2.51% and 15.53% test error, respectively, despite searching in a space where each cell has 150 times as many possible configurations than in the DARTS baseline. Code is available at https://github.com/cogsys-tuebingen/prdarts

Peizheng Li, Zhenghao Zhang, David Holtz et al.

End-to-end autonomous driving methods built on vision language models (VLMs) have undergone rapid development driven by their universal visual understanding and strong reasoning capabilities obtained from the large-scale pretraining. However, we find that current VLMs struggle to understand fine-grained 3D spatial relationships which is a fundamental requirement for systems interacting with the physical world. To address this issue, we propose SpaceDrive, a spatial-aware VLM-based driving framework that treats spatial information as explicit positional encodings (PEs) instead of textual digit tokens, enabling joint reasoning over semantic and spatial representations. SpaceDrive employs a universal positional encoder to all 3D coordinates derived from multi-view depth estimation, historical ego-states, and text prompts. These 3D PEs are first superimposed to augment the corresponding 2D visual tokens. Meanwhile, they serve as a task-agnostic coordinate representation, replacing the digit-wise numerical tokens as both inputs and outputs for the VLM. This mechanism enables the model to better index specific visual semantics in spatial reasoning and directly regress trajectory coordinates rather than generating digit-by-digit, thereby enhancing planning accuracy. Extensive experiments validate that SpaceDrive achieves state-of-the-art open-loop performance on the nuScenes dataset and the second-best Driving Score of 78.02 on the Bench2Drive closed-loop benchmark over existing VLM-based methods.

Marcel Hallgarten, Julian Zapata, Martin Stoll et al.

Real-world autonomous driving systems must make safe decisions in the face of rare and diverse traffic scenarios. Current state-of-the-art planners are mostly evaluated on real-world datasets like nuScenes (open-loop) or nuPlan (closed-loop). In particular, nuPlan seems to be an expressive evaluation method since it is based on real-world data and closed-loop, yet it mostly covers basic driving scenarios. This makes it difficult to judge a planner's capabilities to generalize to rarely-seen situations. Therefore, we propose a novel closed-loop benchmark interPlan containing several edge cases and challenging driving scenarios. We assess existing state-of-the-art planners on our benchmark and show that neither rule-based nor learning-based planners can safely navigate the interPlan scenarios. A recently evolving direction is the usage of foundation models like large language models (LLM) to handle generalization. We evaluate an LLM-only planner and introduce a novel hybrid planner that combines an LLM-based behavior planner with a rule-based motion planner that achieves state-of-the-art performance on our benchmark.

Thomas Gossard, Julian Krismer, Andreas Ziegler et al.

Spin plays a pivotal role in ball-based sports. Estimating spin becomes a key skill due to its impact on the ball's trajectory and bouncing behavior. Spin cannot be observed directly, making it inherently challenging to estimate. In table tennis, the combination of high velocity and spin renders traditional low frame rate cameras inadequate for quickly and accurately observing the ball's logo to estimate the spin due to the motion blur. Event cameras do not suffer as much from motion blur, thanks to their high temporal resolution. Moreover, the sparse nature of the event stream solves communication bandwidth limitations many frame cameras face. To the best of our knowledge, we present the first method for table tennis spin estimation using an event camera. We use ordinal time surfaces to track the ball and then isolate the events generated by the logo on the ball. Optical flow is then estimated from the extracted events to infer the ball's spin. We achieved a spin magnitude mean error of $10.7 \pm 17.3$ rps and a spin axis mean error of $32.9 \pm 38.2°$ in real time for a flying ball.

Max Beffert, Andreas Zell

One of the main limitations of multirotor UAVs is their short flight time due to battery constraints. A practical solution for continuous operation is to power the drone from the ground via a tether. While this approach has been demonstrated for stationary systems, scenarios with a fast-moving base vehicle or strong wind conditions require modeling the tether forces, including aerodynamic effects. In this work, we propose two complementary approaches for real-time quasi-static tether modeling with aerodynamics. The first is an analytical method based on catenary theory with a uniform drag assumption, achieving very fast solve times below 1~ms. The second is a numerical method that discretizes the tether into segments and lumped masses, solving the equilibrium equations using CasADi and IPOPT. By leveraging initialization strategies, such as warm starting and analytical initialization, real-time performance was achieved with a solve time of 5~ms, while allowing for flexible force formulations. Both approaches were validated in real-world tests using a load cell to measure the tether force. The results show that the analytical method provides sufficient accuracy for most tethered UAV applications with minimal computational cost, while the numerical method offers higher flexibility and physical accuracy when required. These approaches form a lightweight and extensible framework for real-time tether simulation, applicable to both offline optimization and online tasks such as simulation, control, and trajectory planning.

Yuzhi Lai, Shenghai Yuan, Peizheng Li et al.

Gaze is a valuable means of communication for impaired people with extremely limited motor capabilities. However, robust gaze-based intent recognition in multi-object environments is challenging due to gaze noise, micro-saccades, viewpoint changes, and dynamic objects. To address this, we propose an object-centric gaze grounding framework that stabilizes intent through a sticky-glance algorithm, jointly modeling geometric distance and direction trends. The inferred intent remains anchored to the object even under short glances with minimal 3 gaze samples, achieving a tracking rate of 0.94 for dynamic targets and selection accuracy of 0.98 for static targets. We further introduce a continuous shared control and multi-modal interaction paradigm, enabling high-readiness control and human-in-loop feedback, thereby reducing task duration for nearly 10 \%. Experiments across dynamic tracking, multi-perspective alignment, a baseline comparison, user studies, and ablation studies demonstrate improved robustness, efficiency, and reduced workload compared to representative baselines.

Andreas Ziegler, Karl Vetter, Thomas Gossard et al.

Neuromorphic Computing (NC) and Spiking Neural Networks (SNNs) in particular are often viewed as the next generation of Neural Networks (NNs). NC is a novel bio-inspired paradigm for energy efficient neural computation, often relying on SNNs in which neurons communicate via spikes in a sparse, event-based manner. This communication via spikes can be exploited by neuromorphic hardware implementations very effectively and results in a drastic reductions of power consumption and latency in contrast to regular GPU-based NNs. In recent years, neuromorphic hardware has become more accessible, and the support of learning frameworks has improved. However, available hardware is partially still experimental, and it is not transparent what these solutions are effectively capable of, how they integrate into real-world robotics applications, and how they realistically benefit energy efficiency and latency. In this work, we provide the robotics research community with an overview of what is possible with SNNs on neuromorphic hardware focusing on real-time processing. We introduce a benchmark of three popular neuromorphic hardware devices for the task of event-based object detection. Moreover, we show that an SNN on a neuromorphic hardware is able to run in a challenging table tennis robot setup in real-time.

Andreas Ziegler, Thomas Gossard, Arren Glover et al.

Table tennis robots gained traction over the last years and have become a popular research challenge for control and perception algorithms. Fast and accurate ball detection is crucial for enabling a robotic arm to rally the ball back successfully. So far, most table tennis robots use conventional, frame-based cameras for the perception pipeline. However, frame-based cameras suffer from motion blur if the frame rate is not high enough for fast-moving objects. Event-based cameras, on the other hand, do not have this drawback since pixels report changes in intensity asynchronously and independently, leading to an event stream with a temporal resolution on the order of us. To the best of our knowledge, we present the first real-time perception pipeline for a table tennis robot that uses only event-based cameras. We show that compared to a frame-based pipeline, event-based perception pipelines have an update rate which is an order of magnitude higher. This is beneficial for the estimation and prediction of the ball's position, velocity, and spin, resulting in lower mean errors and uncertainties. These improvements are an advantage for the robot control, which has to be fast, given the short time a table tennis ball is flying until the robot has to hit back.

Benjamin Kiefer, Yitong Quan, Andreas Zell

Unmanned surface vehicles (USVs) and boats are increasingly important in maritime operations, yet their deployment is limited due to costly sensors and complexity. LiDAR, radar, and depth cameras are either costly, yield sparse point clouds or are noisy, and require extensive calibration. Here, we introduce a novel approach for approximate distance estimation in USVs using supervised object detection. We collected a dataset comprising images with manually annotated bounding boxes and corresponding distance measurements. Leveraging this data, we propose a specialized branch of an object detection model, not only to detect objects but also to predict their distances from the USV. This method offers a cost-efficient and intuitive alternative to conventional distance measurement techniques, aligning more closely with human estimation capabilities. We demonstrate its application in a marine assistance system that alerts operators to nearby objects such as boats, buoys, or other waterborne hazards.

Rafia Rahim, Samuel Woerz, Andreas Zell

Recently, end-to-end deep networks based stereo matching methods, mainly because of their performance, have gained popularity. However, this improvement in performance comes at the cost of increased computational and memory bandwidth requirements, thus necessitating specialized hardware (GPUs); even then, these methods have large inference times compared to classical methods. This limits their applicability in real-world applications. Although we desire high accuracy stereo methods albeit with reasonable inference time. To this end, we propose a fast end-to-end stereo matching method. Majority of this speedup comes from integrating a leaner backbone. To recover the performance lost because of a leaner backbone, we propose to use learned attention weights based cost volume combined with LogL1 loss for stereo matching. Using LogL1 loss not only improves the overall performance of the proposed network but also leads to faster convergence. We do a detailed empirical evaluation of different design choices and show that our method requires 4x less operations and is also about 9 to 14x faster compared to the state of the art methods like ACVNet [1], LEAStereo [2] and CFNet [3] while giving comparable performance.

Thomas Gossard, Andreas Ziegler, Andreas Zell

Sports analysis requires processing large amounts of data, which is time-consuming and costly. Advancements in neural networks have significantly alleviated this burden, enabling highly accurate ball tracking in sports broadcasts. However, relying solely on 2D ball tracking is limiting, as it depends on the camera's viewpoint and falls short of supporting comprehensive game analysis. To address this limitation, we propose a novel approach for reconstructing precise 3D ball trajectories from online table tennis match recordings. Our method leverages the underlying physics of the ball's motion to identify the bounce state that minimizes the reprojection error of the ball's flying trajectory, hence ensuring an accurate and reliable 3D reconstruction. A key advantage of our approach is its ability to infer ball spin without relying on human pose estimation or racket tracking, which are often unreliable or unavailable in broadcast footage. We developed an automated camera calibration method capable of reliably tracking camera movements. Additionally, we adapted an existing 3D pose estimation model, which lacks depth motion capture, to accurately track player movements. Together, these contributions enable the full 3D reconstruction of a table tennis rally.

Yitong Quan, Benjamin Kiefer, Martin Messmer et al.

Accurate distance estimation is a fundamental challenge in robotic perception, particularly in omnidirectional imaging, where traditional geometric methods struggle with lens distortions and environmental variability. In this work, we propose a neural network-based approach for monocular distance estimation using a single 360° fisheye lens camera. Unlike classical trigonometric techniques that rely on precise lens calibration, our method directly learns and infers the distance of objects from raw omnidirectional inputs, offering greater robustness and adaptability across diverse conditions. We evaluate our approach on three 360° datasets (LOAF, ULM360, and a newly captured dataset Boat360), each representing distinct environmental and sensor setups. Our experimental results demonstrate that the proposed learning-based model outperforms traditional geometry-based methods and other learning baselines in both accuracy and robustness. These findings highlight the potential of deep learning for real-time omnidirectional distance estimation, making our approach particularly well-suited for low-cost applications in robotics, autonomous navigation, and surveillance.

Yitong Quan, Benjamin Kiefer, Martin Messmer et al.

Modern image-based object detection models, such as YOLOv7, primarily process individual frames independently, thus ignoring valuable temporal context naturally present in videos. Meanwhile, existing video-based detection methods often introduce complex temporal modules, significantly increasing model size and computational complexity. In practical applications such as surveillance and autonomous driving, transient challenges including motion blur, occlusions, and abrupt appearance changes can severely degrade single-frame detection performance. To address these issues, we propose a straightforward yet highly effective strategy: stacking multiple consecutive frames as input to a YOLO-based detector while supervising only the output corresponding to a single target frame. This approach leverages temporal information with minimal modifications to existing architectures, preserving simplicity, computational efficiency, and real-time inference capability. Extensive experiments on the challenging MOT20Det and our BOAT360 datasets demonstrate that our method improves detection robustness, especially for lightweight models, effectively narrowing the gap between compact and heavy detection networks. Additionally, we contribute the BOAT360 benchmark dataset, comprising annotated fisheye video sequences captured from a boat, to support future research in multi-frame video object detection in challenging real-world scenarios.

Andreas Ziegler, David Joseph, Thomas Gossard et al.

Event-based cameras are bio-inspired sensors that detect light changes asynchronously for each pixel. They are increasingly used in fields like computer vision and robotics because of several advantages over traditional frame-based cameras, such as high temporal resolution, low latency, and high dynamic range. As with any camera, the output's quality depends on how well the camera's settings, called biases for event-based cameras, are configured. While frame-based cameras have advanced automatic configuration algorithms, there are very few such tools for tuning these biases. A systematic testing framework would require observing the same scene with different biases, which is tricky since event cameras only generate events when there is movement. Event simulators exist, but since biases heavily depend on the electrical circuit and the pixel design, available simulators are not well suited for bias tuning. To allow reproducibility, we present BiasBench, a novel event dataset containing multiple scenes with settings sampled in a grid-like pattern. We present three different scenes, each with a quality metric of the downstream application. Additionally, we present a novel, RL-based method to facilitate online bias adjustments.

Rafia Rahim, Samuel Woerz, Andreas Zell

Knowledge distillation has been quite popular in vision for tasks like classification and segmentation however not much work has been done for distilling state-of-the-art stereo matching methods despite their range of applications. One of the reasons for its lack of use in stereo matching networks is due to the inherent complexity of these networks, where a typical network is composed of multiple two- and three-dimensional modules. In this work, we systematically combine the insights from state-of-the-art stereo methods with general knowledge-distillation techniques to develop a joint framework for stereo networks distillation with competitive results and faster inference. Moreover, we show, via a detailed empirical analysis, that distilling knowledge from the stereo network requires careful design of the complete distillation pipeline starting from backbone to the right selection of distillation points and corresponding loss functions. This results in the student networks that are not only leaner and faster but give excellent performance . For instance, our student network while performing better than the performance oriented methods like PSMNet [1], CFNet [2], and LEAStereo [3]) on benchmark SceneFlow dataset, is 8x, 5x, and 8x faster respectively. Furthermore, compared to speed oriented methods having inference time less than 100ms, our student networks perform better than all the tested methods. In addition, our student network also shows better generalization capabilities when tested on unseen datasets like ETH3D and Middlebury.

Hamd ul Moqeet Riaz, Nuri Benbarka, Timon Hoefer et al.

We present FourierMask, which employs Fourier series combined with implicit neural representations to generate instance segmentation masks. We apply a Fourier mapping (FM) to the coordinate locations and utilize the mapped features as inputs to an implicit representation (coordinate-based multi-layer perceptron (MLP)). FourierMask learns to predict the coefficients of the FM for a particular instance, and therefore adapts the FM to a specific object. This allows FourierMask to be generalized to predict instance segmentation masks from natural images. Since implicit functions are continuous in the domain of input coordinates, we illustrate that by sub-sampling the input pixel coordinates, we can generate higher resolution masks during inference. Furthermore, we train a renderer MLP (FourierRend) on the uncertain predictions of FourierMask and illustrate that it significantly improves the quality of the masks. FourierMask shows competitive results on the MS COCO dataset compared to the baseline Mask R-CNN at the same output resolution and surpasses it on higher resolution.

Yapeng Gao, Jonas Tebbe, Andreas Zell

Learning to play table tennis is a challenging task for robots, as a wide variety of strokes required. Recent advances have shown that deep Reinforcement Learning (RL) is able to successfully learn the optimal actions in a simulated environment. However, the applicability of RL in real scenarios remains limited due to the high exploration effort. In this work, we propose a realistic simulation environment in which multiple models are built for the dynamics of the ball and the kinematics of the robot. Instead of training an end-to-end RL model, a novel policy gradient approach with TD3 backbone is proposed to learn the racket strokes based on the predicted state of the ball at the hitting time. In the experiments, we show that the proposed approach significantly outperforms the existing RL methods in simulation. Furthermore, to cross the domain from simulation to reality, we adopt an efficient retraining method and test it in three real scenarios. The resulting success rate is 98% and the distance error is around 24.9 cm. The total training time is about 1.5 hours.

Nuri Benbarka, Timon Höfer, Hamd ul-moqeet Riaz et al.

Implicit Neural Representations (INR) use multilayer perceptrons to represent high-frequency functions in low-dimensional problem domains. Recently these representations achieved state-of-the-art results on tasks related to complex 3D objects and scenes. A core problem is the representation of highly detailed signals, which is tackled using networks with periodic activation functions (SIRENs) or applying Fourier mappings to the input. This work analyzes the connection between the two methods and shows that a Fourier mapped perceptron is structurally like one hidden layer SIREN. Furthermore, we identify the relationship between the previously proposed Fourier mapping and the general d-dimensional Fourier series, leading to an integer lattice mapping. Moreover, we modify a progressive training strategy to work on arbitrary Fourier mappings and show that it improves the generalization of the interpolation task. Lastly, we compare the different mappings on the image regression and novel view synthesis tasks. We confirm the previous finding that the main contributor to the mapping performance is the size of the embedding and standard deviation of its elements.

Maximus Mutschler, Kevin Laube, Andreas Zell

A fundamental challenge in Deep Learning is to find optimal step sizes for stochastic gradient descent automatically. In traditional optimization, line searches are a commonly used method to determine step sizes. One problem in Deep Learning is that finding appropriate step sizes on the full-batch loss is unfeasibly expensive. Therefore, classical line search approaches, designed for losses without inherent noise, are usually not applicable. Recent empirical findings suggest, inter alia, that the full-batch loss behaves locally parabolically in the direction of noisy update step directions. Furthermore, the trend of the optimal update step size changes slowly. By exploiting these and more findings, this work introduces a line-search method that approximates the full-batch loss with a parabola estimated over several mini-batches. Learning rates are derived from such parabolas during training. In the experiments conducted, our approach is on par with SGD with Momentum tuned with a piece-wise constant learning rate schedule and often outperforms other line search approaches for Deep Learning across models, datasets, and batch sizes on validation and test accuracy. In addition, our approach is the first line search approach for Deep Learning that samples a larger batch size over multiple inferences to still work in low-batch scenarios.

Rafia Rahim, Faranak Shamsafar, Andreas Zell

Deep learning based 3D stereo networks give superior performance compared to 2D networks and conventional stereo methods. However, this improvement in the performance comes at the cost of increased computational complexity, thus making these networks non-practical for the real-world applications. Specifically, these networks use 3D convolutions as a major work horse to refine and regress disparities. In this work first, we show that these 3D convolutions in stereo networks consume up to 94% of overall network operations and act as a major bottleneck. Next, we propose a set of "plug-&-run" separable convolutions to reduce the number of parameters and operations. When integrated with the existing state of the art stereo networks, these convolutions lead up to 7x reduction in number of operations and up to 3.5x reduction in parameters without compromising their performance. In fact these convolutions lead to improvement in their performance in the majority of cases.

Timon Höfer, Faranak Shamsafar, Nuri Benbarka et al.

Bin picking is a core problem in industrial environments and robotics, with its main module as 6D pose estimation. However, industrial depth sensors have a lack of accuracy when it comes to small objects. Therefore, we propose a framework for pose estimation in highly cluttered scenes with small objects, which mainly relies on RGB data and makes use of depth information only for pose refinement. In this work, we compare synthetic data generation approaches for object detection and pose estimation and introduce a pose filtering algorithm that determines the most accurate estimated poses. We will make our

Leon Amadeus Varga, Andreas Zell

Object detection on Unmanned Aerial Vehicles (UAVs) is still a challenging task. The recordings are mostly sparse and contain only small objects. In this work, we propose a simple tiling method that improves the detection capability in the remote sensing case without modifying the model itself. By reducing the background bias and enabling the usage of higher image resolutions during training, our method can improve the performance of models substantially. The procedure was validated on three different data sets and outperformed similar approaches in performance and speed.

Leon Amadeus Varga, Benjamin Kiefer, Martin Messmer et al.

Unmanned Aerial Vehicles (UAVs) are of crucial importance in search and rescue missions in maritime environments due to their flexible and fast operation capabilities. Modern computer vision algorithms are of great interest in aiding such missions. However, they are dependent on large amounts of real-case training data from UAVs, which is only available for traffic scenarios on land. Moreover, current object detection and tracking data sets only provide limited environmental information or none at all, neglecting a valuable source of information. Therefore, this paper introduces a large-scaled visual object detection and tracking benchmark (SeaDronesSee) aiming to bridge the gap from land-based vision systems to sea-based ones. We collect and annotate over 54,000 frames with 400,000 instances captured from various altitudes and viewing angles ranging from 5 to 260 meters and 0 to 90 degrees while providing the respective meta information for altitude, viewing angle and other meta data. We evaluate multiple state-of-the-art computer vision algorithms on this newly established benchmark serving as baselines. We provide an evaluation server where researchers can upload their prediction and compare their results on a central leaderboard

Kevin Alexander Laube, Andreas Zell

Modern Neural Architecture Search methods have repeatedly broken state-of-the-art results for several disciplines. The super-network, a central component of many such methods, enables quick estimates of accuracy or loss statistics for any architecture in the search space. They incorporate the network weights of all candidate architectures and can thus approximate specific ones by applying the respective operations. However, this design ignores potential dependencies between consecutive operations. We extend super-networks with conditional weights that depend on combinations of choices and analyze their effect. Experiments in NAS-Bench 201 and NAS-Bench-Macro-based search spaces show improvements in the architecture selection and that the resource overhead is nearly negligible for sequential network designs.

Leon Amadeus Varga, Jan Makowski, Andreas Zell

We present a system to measure the ripeness of fruit with a hyperspectral camera and a suitable deep neural network architecture. This architecture did outperform competitive baseline models on the prediction of the ripeness state of fruit. For this, we recorded a data set of ripening avocados and kiwis, which we make public. We also describe the process of data collection in a manner that the adaption for other fruit is easy. The trained network is validated empirically, and we investigate the trained features. Furthermore, a technique is introduced to visualize the ripening process.

Maximus Mutschler, Andreas Zell

Optimization in Deep Learning is mainly guided by vague intuitions and strong assumptions, with a limited understanding how and why these work in practice. To shed more light on this, our work provides some deeper understandings of how SGD behaves by empirically analyzing the trajectory taken by SGD from a line search perspective. Specifically, a costly quantitative analysis of the full-batch loss along SGD trajectories from common used models trained on a subset of CIFAR-10 is performed. Our core results include that the full-batch loss along lines in update step direction is highly parabolically. Further on, we show that there exists a learning rate with which SGD always performs almost exact line searches on the full-batch loss. Finally, we provide a different perspective why increasing the batch size has almost the same effect as decreasing the learning rate by the same factor.

Martin Messmer, Benjamin Kiefer, Andreas Zell

This work introduces a new preprocessing step for object detection applicable to UAV bird's eye view imagery, which we call Adaptive Resizing. By design, it helps alleviate the challenges coming with the vast variances in objects' scales, naturally inherent to UAV data sets. Furthermore, it improves inference speed by two to three times on average. We test this extensively on UAVDT, VisDrone, and on a new data set we captured ourselves and achieve consistent improvements while being considerably faster. Moreover, we show how to apply this method to generic UAV object detection tasks. Additionally, we successfully test our approach on a height transfer task where we train on some interval of altitudes and test on a different one. Furthermore, we introduce a small, fast detector meant for deployment to an embedded GPU. Code will be made publicly available on our website.

Benjamin Kiefer, Martin Messmer, Andreas Zell

Object detection from Unmanned Aerial Vehicles (UAVs) is of great importance in many aerial vision-based applications. Despite the great success of generic object detection methods, a significant performance drop is observed when applied to images captured by UAVs. This is due to large variations in imaging conditions, such as varying altitudes, dynamically changing viewing angles, and different capture times. These variations lead to domain imbalances and, thus, trained models suffering from domain bias. We demonstrate that domain knowledge is a valuable source of information and thus propose domain-aware object detectors by using freely accessible sensor data. By splitting the model into cross-domain and domain-specific parts, substantial performance improvements are achieved on multiple data sets across various models and metrics without changing the architecture. In particular, we achieve a new state-of-the-art performance on UAVDT for embedded real-time detectors. Furthermore, we create a new airborne image data set by annotating 13,713 objects in 2,900 images featuring precise altitude and viewing angle annotations.

Jonas Tebbe, Lukas Krauch, Yapeng Gao et al.

Reinforcement learning (RL) has achieved some impressive recent successes in various computer games and simulations. Most of these successes are based on having large numbers of episodes from which the agent can learn. In typical robotic applications, however, the number of feasible attempts is very limited. In this paper we present a sample-efficient RL algorithm applied to the example of a table tennis robot. In table tennis every stroke is different, with varying placement, speed and spin. An accurate return therefore has to be found depending on a high-dimensional continuous state space. To make learning in few trials possible the method is embedded into our robot system. In this way we can use a one-step environment. The state space depends on the ball at hitting time (position, velocity, spin) and the action is the racket state (orientation, velocity) at hitting. An actor-critic based deterministic policy gradient algorithm was developed for accelerated learning. Our approach performs competitively both in a simulation and on the real robot in a number of challenging scenarios. Accurate results are obtained without pre-training in under $200$ episodes of training. The video presenting our experiments is available at https://youtu.be/uRAtdoL6Wpw.

Maximus Mutschler, Andreas Zell

A fundamental challenge in deep learning is that the optimal step sizes for update steps of stochastic gradient descent are unknown. In traditional optimization, line searches are used to determine good step sizes, however, in deep learning, it is too costly to search for good step sizes on the expected empirical loss due to noisy losses. This empirical work shows that it is possible to approximate the expected empirical loss on vertical cross sections for common deep learning tasks considerably cheaply. This is achieved by applying traditional one-dimensional function fitting to measured noisy losses of such cross sections. The step to a minimum of the resulting approximation is then used as step size for the optimization. This approach leads to a robust and straightforward optimization method which performs well across datasets and architectures without the need of hyperparameter tuning.