Repository for Publications and Research Data
News from the ETH Library
Recently Added
Efficient Data Query Execution as Serverless Workflows
Item type: Master Thesis
Stocker, Tobias (2025)
Serverless computing, and Function-as-a-Service (FaaS) in particular, has attracted a lot of attention from research and industry in recent years. It allows users to launch thousands of short-lived functions and only bills them for the services actively used, enabling highly elastic execution. Following the success of serverless computing, new systems have been proposed to optimize various efficiency shortcomings of existing FaaS platforms. These systems include Dirigent, a new cluster manager designed for serverless orchestration, and Dandelion, a new sandbox runtime, proposing a new programming model for serverless applications, which requires applications to run as compositions of functions. At the same time, many new data processing frameworks have been proposed to benefit from the high elasticity offered by serverless platforms to improve efficiency and reduce cost. In this work, we enable Dirigent to orchestrate the execution of Dandelion workflows over a cluster of Dandelion worker nodes. Using this system, we implement a data query execution system that runs data queries as Dandelion workflows with operators as serverless functions. We conduct a range of experiments and show the efficiency benefits of our system coming from the Dandelion programming model by running the Star Schema Benchmarks queries with gigabytes of data among other workloads. Using roughly 700 MB of input data and a single Dandelion worker, we achieve SSB query latencies ranging from 1.9 to 3.5 seconds, outperforming DuckDB in a comparable setup where data needs to be fetched from remote storage. Using 10 workers, we achieve SSB query latencies ranging between 7.5 and 10 seconds processing 6.9 GB of data.
Bridging the Real and Virtual Worlds: Neural Advances toward Realistic Dynamic Digital Models
Item type: Doctoral Thesis
Mihajlovic, Marko (2025)
Imagine a technology capable of creating and simulating digital environments so immersive and precise that they are nearly indistinguishable from the physical world. Such advancements could revolutionize scientific research, enabling complex systems to be modeled and safely explored in controlled conditions, with applications across fields like biology and physics.
Realizing this vision requires interactive digital models that allow humans to engage with detailed, dynamic replicas of physical systems. This dissertation addresses several critical challenges toward creating fully immersive simulations.
First, as humans are often the central element in digital environments, this work begins by examining algorithms that learn neural representations to model 3D humans, facilitating realistic human-scene interactions within complex 3D environments.
Traditional approaches rely on precomputed distance fields from scans to prevent collisions, which are error-prone and require high-quality scans.
This dissertation presents a novel solution by modeling the human body as a continuous volumetric field, enabling efficient, differentiable collision handling. The proposed model, LEAP, is the first neural volumetric body model, defined by a deformation field and neural isosurface that represent human shapes and poses, seamlessly integrating with optimization frameworks.
Building on LEAP, the more advanced COAP model improves robustness by incorporating kinematic and structural priors, allowing it to accurately handle a broader range of human poses and shapes.
Such volumetric neural body models offer a flexible and robust solution for achieving realistic human interactions within digital scenes.
The second focus of this dissertation addresses challenges in the data capture process behind the realistic digital models. Current reconstruction techniques often require hundreds of carefully captured images to ensure high-quality 3D models, making the process time-consuming and limiting its accessibility. To broaden access to this technology, this dissertation introduces KeypointNeRF, a novel Neural Radiance Field (NeRF) model capable of reconstructing high-fidelity digital humans from as few as 2-3 images. KeypointNeRF leverages human keypoints to learn consistent features across diverse datasets, effectively capturing local human details and generalizing well to unseen individuals, enabling users to create immersive digital models with minimal input.
Recognizing KeypointNeRF's limitation in handling generic dynamic scenes, this dissertation introduces ResFields, a novel multi-layer perceptron (MLP) architecture optimized for reconstructing spatiotemporal signals from multi-view video streams. ResFields replaces static MLP layers with time-dependent layers, using trainable residual parameters to boost learning capacity while maintaining a compact architecture, minimizing runtime overhead. This architecture is versatile and compatible with most MLP-based methods, making it a powerful tool for reconstructing dynamic digital scenes in the combination with dynamic NeRF models.
Finally, as the ResField model relies on NeRF for modeling dynamic scenes, it inherits NeRF's limitations like slow rendering speed and training time.
Therefore, the last part of this dissertation introduces a real-time neural Gaussian splatting (3DGS) model, named SplatFields, for high-fidelity reconstruction from limited input views. It supports both static and dynamic reconstructions by incorporating learnable neural elements that enhance spatial autocorrelation among Gaussian splats. The ResField architecture is further utilized to scale this model for dynamic sequences, achieving state-of-the-art performance without reliance on external data-driven priors.
By advancing these technologies, this dissertation brings us closer to the vision of fully realistic and interactive digital simulations.
The Magic Box: Stabilizing an Inverted Pendulum with Hidden Infrared Vision
Item type: Master Thesis
Czubarow, Anthony (2025)
The vision behind this work is the creation of a Magic Box : a platform on which everyday objects, such as pens, can stand upright and appear to balance as if by magic, with sensing and actuation completely hidden from the observer. As a first step toward this goal, the thesis presents a proof of concept based on a one-dimensional inverted pendulum stabilized using a time-of-flight (ToF) camera concealed beneath an infrared-pass filter. This setting introduces significant challenges: the ToF camera provides noisy measurements and is restricted to a bottom-up view, yet the system requires accurate angle estimates in real time for stabilization. Addressing these challenges extends the long tradition of inverted pendulum research, where stabilization has typically depended on classical sensors such as encoders or external cameras. A discrete-time Linear Quadratic Gaussian (LQG) regulator was designed for the linearized cart-pendulum model. The state estimator fused measurements from the motor’s Hall-e!ect sensor and the ToF camera to provide the feedback required for control. The physical damping parameters were identified via batch constrained non-linear least squares fitting from experimental trajectories. The system parameters needed for the Kalman filter were tuned automatically through Bayesian optimization against ground-truth angles from an external camera. All sensing, actuation, and computation were integrated into a compact prototype with a custom carriage, guideway, rack-and-pinion drive, and a Raspberry Pi, designed to be easily reproducible and open-source. Experimental results demonstrate reliable stabilization across repeated trials, validating the feasibility of the Magic Box.
Scale, Structure, and Bias: Interplay and Trade-Offs in Deep Learning
Item type: Doctoral Thesis
Anagnostidis , Sotirios Konstantinos (2025)
Development and Evaluation of Hourly Temperature Analysis Data Sets in Complex Terrain
Item type: Doctoral Thesis
Frey, Louis (2025)