A5031102089- Autonomous Vehicle Technology and Safety
- Medical Imaging and Analysis
- Video Surveillance and Tracking Methods
- Advanced Neural Network Applications
- Advanced X-ray and CT Imaging
- Medical Imaging Techniques and Applications
- Pelvic and Acetabular Injuries
- Domain Adaptation and Few-Shot Learning
- Quantum Computing Algorithms and Architecture
- Multimodal Machine Learning Applications
- Quantum Mechanics and Applications
- Human Pose and Action Recognition
- Anatomy and Medical Technology
- Quantum Information and Cryptography
- Radiation Dose and Imaging
- Geophysical Methods and Applications
- Neural Networks and Applications
- COVID-19 diagnosis using AI
- Radiomics and Machine Learning in Medical Imaging
- Adversarial Robustness in Machine Learning
- Seismic Imaging and Inversion Techniques
- Dental Radiography and Imaging
- Simulation Techniques and Applications
- Air Quality Monitoring and Forecasting
- Generative Adversarial Networks and Image Synthesis
ETH Zurich
2019-2024
Sofia University "St. Kliment Ohridski"
2024
Johns Hopkins University
2018-2020
Friedrich-Alexander-Universität Erlangen-Nürnberg
2018-2019
Board of the Swiss Federal Institutes of Technology
2019
Autonomous systems that operate in dynamic environments require robust object tracking 3D as one of their key components. Most recent approaches for multi-object (MOT) from LIDAR use dynamics together with a set handcrafted features to match detections objects across multiple frames. However, manually designing such and heuristics is cumbersome often leads suboptimal performance. In this work, we instead strive towards unified learning based approach the MOT problem. We design graph...
We study the problem of robust domain adaptation in context unavailable target labels and source data. The considered robustness is against adversarial perturbations. This paper aims at answering question finding right strategy to make model accurate setting unsupervised without major findings this are: (i) models can be transferred robustly target; (ii) greatly benefit from nonrobust pseudo-labels pair-wise contrastive loss. proposed method using non-robust performs surprisingly well on...
Multi-Object Tracking (MOT) is most often approached in the tracking-by-detection paradigm, where object detections are associated through time. The association step naturally leads to discrete optimization problems. As these problems NP-hard, they can only be solved exactly for small instances on current hardware. Adiabatic quantum computing (AQC) offers a solution this, as it has potential provide considerable speedup range of NP-hard near future. However, MOT formulations unsuitable due...
Differentiable simulators have recently shown great promise for training autonomous vehicle controllers. Being able to backpropagate through them, they can be placed into an end-to-end loop where their known dynamics turn useful priors the policy learn, removing typical black box assumption of environment. So far, these systems only been used train policies. However, this is not end story in terms what offer. Here, first time, we use them world models. Specifically, present three new task...
This work addresses the problem of semantic scene understanding under foggy road conditions. Although marked progress has been made in over recent years, it is mainly concentrated on clear weather outdoor scenes. Extending segmentation methods to adverse conditions like fog crucially important for applications such as self-driving cars. In this paper, we propose a novel method, which uses purely synthetic data improve performance unseen real-world scenes captured streets Zurich and its...
With the advent of robotic C-arm computed tomography (CT) systems in medicine and twin-robotic CT industry, new possibilities for realisation complex trajectories scans are emerging. These will increase range applications, enable optimisation image quality many applications open up to reduce scan time radiation dose. In this work, trajectory methods optimising both, task-based data completeness, presented by combining two different metrics. On one hand, is optimised with a proven observer...
This work studies the problem of predicting sequence future actions for surrounding vehicles in real-world driving scenarios. To this aim, we make three main contributions. The first contribution is an automatic method to convert trajectories recorded scenarios action sequences with help HD maps. enables dataset creation task from large-scale data. Our second lies applying well-known traffic agent tracking and prediction Argoverse, resulting 228,000 sequences. Additionally, 2,245 were...
Forecasting the future behavior of all traffic agents in vicinity is a key task to achieve safe and reliable autonomous driving systems. It challenging problem as adjust their depending on intentions, others' actions, road layout. In this paper, we propose Decoder Fusion RNN (DF-RNN), recurrent, attention-based approach for motion forecasting. Our network composed recurrent encoder, an inter-agent multi-headed attention module, context-aware decoder. We design map encoder that embeds...
Comprehensive semantic segmentation is one of the key components for robust scene understanding and a requirement to enable autonomous driving. Driven by large scale datasets, convolutional neural networks show impressive results on this task. However, algorithm generalizing various scenes conditions would require an enormously diverse dataset, making labour intensive data acquisition labeling process prohibitively expensive. Under assumption structural similarities between maps, domain...
Monitoring a fleet of robots requires stable long-term tracking with re-identification, which is yet an unsolved challenge in many scenarios. One application this the analysis autonomous robotic soccer games at RoboCup. Tracking these handling identically looking players, strong occlusions, and non-professional video recordings, but also offers state information estimated by robots. In order to make effective use coming from robot sensors, we propose robust identification pipeline. It fuses...
Recent progress in large language models and access to large-scale robotic datasets has sparked a paradigm shift robotics transforming them into generalists able adapt various tasks, scenes, robot modalities. A step for the community are open Vision Language Action which showcase strong performance wide variety of tasks. In this work, we study visual generalization capabilities three existing foundation models, propose corresponding evaluation framework. Our shows that do not exhibit...
3D scene understanding is a long-standing challenge in computer vision and key component enabling mixed reality, wearable computing, embodied AI. Providing solution to these applications requires multifaceted approach that covers scene-centric, object-centric, as well interaction-centric capabilities. While there exist numerous datasets approaching the former two problems, task of interactable articulated objects underrepresented only partly covered by current works. In this work, we address...
TL;DR: Gaussian Splatting is a widely adopted approach for 3D scene representation that offers efficient, high-quality reconstruction and rendering. A major reason the success of 3DGS its simplicity representing with set Gaussians, which makes it easy to interpret adapt. To enhance understanding beyond visual representation, approaches have been developed extend semantic vision-language features, especially allowing open-set tasks. In this setting, language features are often aggregated from...
Compressed sensing (CS) deals with the problem of reconstructing a sparse vector from an under-determined set observations. Approximate message passing (AMP) is technique used in CS based on iterative thresholding and inspired by belief propagation graphical models. Due to high transmission rate molecular absorption, spreading loss reflection loss, discrete-time channel impulse response (CIR) typical indoor THz very long exhibits approximately characteristic. In this paper, we develop AMP...