Open Semester Projects and Theses

Low-Latency 4D Reconstruction from Sparse Camera Views in Professional Ice Hockey

Photogrammetry and Remote Sensing (Prof. Schindler)

Volumetric ("4D") reconstruction recovers a dynamic 3D scene from synchronised multi-view video. Gaussian Splatting and generative refinement have made it practical, but it remains slow: per-scene optimisation takes minutes to hours, while most applications need seconds. It also depends on dense camera coverage. Both limits come down to the same quantity, the number of views a method needs at inference time.

Keywords

4D reconstruction, Gaussian Splatting, novel view synthesis, feed-forward reconstruction, sparse views, distillation, sports video

Labels

Semester Project , Master Thesis

Description

Goal

Contact Details

More information

Open this project...

Published since: 2026-08-12 , Earliest start: 2026-09-01 , Latest end: 2027-08-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Metzger Nando

Topics Information, Computing and Communication Sciences

Streaming Scene State for Calibrated Multi-Camera Rigs

Photogrammetry and Remote Sensing (Prof. Schindler)

Feed-forward reconstruction models process a window of frames and produce geometry in one pass, but a live multi-camera stream never ends. Recomputing a full window for every update spends compute on frames already seen, and latency grows with the window. Recent stateful models compress an image collection into a compact hidden state and reconstruct in linear time over hundreds of frames, yet every released system is monocular and essentially static. The open question is what a scene state should look like when the cameras are many, synchronised and calibrated.

Keywords

streaming 3D reconstruction, test-time training, recurrent state, multi-view video, self-distillation, bounded latency

Labels

Master Thesis

Description

Goal

Contact Details

More information

Open this project...

Published since: 2026-08-12 , Earliest start: 2026-09-01 , Latest end: 2027-08-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Metzger Nando

Topics Information, Computing and Communication Sciences

Learned Perceptual Metrics for Occlusion-Heavy Sports Content

Photogrammetry and Remote Sensing (Prof. Schindler)

Reconstruction and generative refinement are judged, in practice, by whether people find the output convincing. Standard metrics disagree with people exactly where it matters: PSNR rewards blur, SSIM is dominated by static background, and full-frame averages hide localised failures. In sports those failures concentrate in occlusion-heavy moments with many similar, fast-moving people, which are also the moments viewers care most about. Blind pairwise human studies answer the question correctly but cost days per evaluation, which makes them unusable inside a training loop.

Keywords

perceptual quality metrics, human preference learning, video quality assessment, novel view synthesis evaluation, sports video

Labels

Master Thesis

Description

Goal

Contact Details

More information

Open this project...

Published since: 2026-08-12 , Earliest start: 2026-11-01 , Latest end: 2026-12-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Metzger Nando

Topics Information, Computing and Communication Sciences

Tree species identification using deep learning

Forest Resources Management

Tree species maps are crucial for effective forest management, biomass assessment, and biodiversity monitoring. Remote sensing products offer flexible and cost-effective ways to assess forest characteristics, while deep learning methods promise high predictive accuracy and transformative applications in forestry. This study aims to apply novel deep learning approaches to detect and identify individual trees and tree species in mixed forests. By addressing the challenges of tree species identification, this research will enhance biodiversity assessment, forest resilience understanding, and management strategies.

Keywords

Tree species identification, computer vision, CNN

Labels

Semester Project , Master Thesis , ETH Zurich (ETHZ)

Description

Goal

Contact Details

More information

Open this project...

Published since: 2026-07-14 , Earliest start: 2026-01-06 , Latest end: 2027-08-31

Applications limited to ETH Zurich , Department of Environmental Systems Science , Department of Civil, Environmental and Geomatic Engineering , Institute of Geodesy and Photogrammetry

Organization Forest Resources Management

Hosts Schindler Konrad , Beloiu Schwenke Mirela , Hangartner Ariane

Topics Agricultural, Veterinary and Environmental Sciences , Information, Computing and Communication Sciences , Engineering and Technology

Multi-View Video Fusion via 3D Geometry

Photogrammetry and Remote Sensing (Prof. Schindler)

Multi-view video fusion—commonly known as video stitching—seeks to merge video streams captured from different physical viewpoints (multi-camera rigs, hand-held cameras, AR/VR setups, autonomous driving stacks) into a single panoramic, temporally coherent video. Existing solutions such as StabStitch++ are fundamentally 2D pipelines: they learn pixel-level warps on the image plane and neglect the underlying 3D scene geometry that should drive cross-view alignment. We propose to treat multi-view video fusion as a genuinely 3D problem, leveraging multi-view geometric foundation models and 3D scene representations (e.g., 3D Gaussian Splatting) to deliver panoramic, temporally stable, low-distortion stitched videos.

Keywords

Video Fusion, Video Stitching, Multi-View, 3D Geometry

Labels

Master Thesis

Description

Video fusion aims to integrate complementary information from paired video sequences (multi-exposure, multi-focus, multi-modal) into a single, temporally coherent video. Recent unified frameworks such as UniVF have shown that exploiting multi-frame spatio-temporal dependencies substantially reduces the flickering and motion discontinuities that plagued earlier frame-by-frame approaches. Yet a fundamental assumption underlies this entire line of work: the input streams are pixel-wise aligned (same viewpoint, same time, different sensors or settings). Once captures originate from different physical viewpoints—multi-camera rigs, hand-held cameras, AR/VR setups, autonomous driving stacks—this assumption breaks down, and the task naturally turns into multi-view video fusion, more commonly known as video stitching. Current video stitching solutions, including the recent unsupervised online framework StabStitch++, address this by learning 2D pixel correspondences and predicting parameterized warps on the image plane, then smoothing the resulting stitching trajectories temporally. While effective for moderate parallax, this paradigm remains fundamentally a 2D reasoning pipeline: it concentrates the alignment burden and projective distortion in the image plane, while neglecting the underlying 3D geometric structure that should guide image stitching. In this project, we propose to go beyond 2D pixel matching and treat multi-view video fusion as a genuinely 3D problem. Concretely, we will: - leverage multi-view geometric foundation models to lift each video stream into a temporally consistent 3D representation and to obtain cross-view correspondences directly in 3D space; - investigate 3D scene representations (e.g., 3D Gaussian Splatting) so that cross-view alignment and seamless blending can be expressed as rendering rather than 2D warping, allowing principled handling of parallax and occlusion; - impose temporal-consistency objectives in the 3D / feature space rather than the photometric plane, so that the stitched video remains coherent across time. The final objective is a 3D-aware multi-view video fusion system that produces panoramic, temporally stable, low-distortion stitched videos.

Goal

Contact Details

More information

Open this project...

Published since: 2026-05-17 , Earliest start: 2026-06-01 , Latest end: 2027-06-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Zhao Zixiang

Topics Information, Computing and Communication Sciences , Engineering and Technology

Bridging Representation and Generation for Multi-Modal Image Fusion

Photogrammetry and Remote Sensing (Prof. Schindler)

Multi-modal image fusion seeks to merge complementary signals from heterogeneous sources (e.g., infrared and visible) into outputs that are simultaneously photometrically faithful and semantically meaningful, but jointly training pixel-level synthesis and high-level semantics end-to-end is data-hungry and tends to degrade the pretrained capabilities of each component. We propose to leverage model stitching to splice a frozen semantic backbone (e.g., DINOv3) with a frozen latent diffusion prior via a lightweight stitching layer, bridging representation and generation for fusion. The resulting system aims to unify pixel fidelity and semantic understanding with minimal task-specific fine-tuning, validated on multi-modal fusion benchmarks together with a downstream perception task.

Keywords

Multi-Modal Image Fusion, Model Stitching, Diffusion Models, Representation Learning

Labels

Master Thesis

Description

Multi-modal (multi-sensor) image fusion aims to merge complementary signals from heterogeneous sources—e.g. infrared and visible—into a single output that is simultaneously photometrically faithful (preserving high-frequency texture and detail) and semantically meaningful (useful for downstream perception). Modern fusion networks, such as CDDFuse and TDFusion, couple the fusion objective with downstream tasks like detection and segmentation so that the fused image is jointly optimized for both visual quality and scene understanding. Yet a tension persists at the heart of this paradigm. Self-supervised representation models such as DINOv3 encode rich, transferable semantics, but their features are too abstract to reconstruct pixel-level texture. Conversely, latent diffusion models are unrivalled at synthesizing high-frequency detail, yet their priors carry little explicit semantic structure. End-to-end joint training of both capabilities is data-hungry and prone to degrading the pretrained capabilities of each component. Recent work on model stitching suggests a far cheaper route: by identifying layers across independently pretrained networks whose activations are related by an (almost) linear map, one can splice the two with a small stitching layer and obtain a combined model that retains the strengths of each side—without large-scale fine-tuning. In this project, we propose to leverage model stitching to bridge representation and generation for multi-modal image fusion. Concretely, we will: - identify, in the spirit of VIST3A, the layer of a frozen diffusion denoiser (DiT) whose activations are most linearly compatible with the features of a strong semantic backbone (e.g. DINOv3), and fit a lightweight stitching projector—initialized in closed form and refined with light SGD (optionally with LoRA on a few DiT layers)—that aligns the two; - couple this stitched representation–generation module with a multi-modal fusion frontend, so that the fused output preserves the high-frequency texture afforded by the diffusion prior while inheriting the semantic structure of the representation backbone; - validate the approach on standard infrared–visible and multi-focus fusion benchmarks together with a downstream perception task (detection or segmentation), benchmarking against CDDFuse and TDFusion. The final objective is a fusion system that unifies pixel fidelity and semantic understanding with minimal task-specific finetuning.

Goal

Contact Details

More information

Open this project...

Published since: 2026-05-17 , Earliest start: 2026-06-01 , Latest end: 2027-06-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Zhao Zixiang

Topics Information, Computing and Communication Sciences , Engineering and Technology

Seeing Seasons: Temporal Self-Supervised Learning from Outdoor Webcam Networks for Forest Monitoring

Photogrammetry and Remote Sensing (Prof. Schindler)

In this project you will build SSL models that learn the language of the seasons from thousands of hours of outdoor webcam footage. By exploiting the interplay between visual and meteorological signals, you will develop representations that capture how forests change over time — and use them to predict ecological events before they happen.

Keywords

Self-Supervised Learning · Temporal Representation Learning · Contrastive Learning · Predictive Architectures · Phenology · Outdoor Webcam Networks · Forest Monitoring · Environmental Computer Vision

Labels

Master Thesis

Description

Phenology is the study of recurring seasonal events in nature, such as when trees first bud in spring or shed their leaves in autumn. These transitions are sensitive indicators of climate change and have direct consequences for forest carbon cycling, biodiversity, and ecosystem management. Operational forest management demands not just observation but early prediction of such events, enabling real-time stress monitoring and timely interventions. Networks of phenocams — high-resolution near-surface cameras capturing tree crowns and canopies every 30 minutes year-round — provide a rich source of such data, with 34 sites across Switzerland and over 700 sites globally. Yet the vast majority of this imagery remains unlabeled, leaving it underexploited beyond simple color-based indices. This thesis explores self-supervised learning (SSL) on these spatio-temporal image sequences, leveraging the interplay between visual and meteorological signals to learn representations that faithfully reflect the phenological trajectory of trees across sites, species, and years — ultimately enabling more accurate and timely phenological monitoring at scale. The central research question of this thesis is: what temporal self-supervised learning objectives can be designed to learn representations that improve phenological monitoring? The intersection of temporal self-supervision and weather as an auxiliary learning signal is a relatively underexplored space, with potential extensions to satellite remote sensing time series. Two promising directions are suggested as potential starting points, though the student is encouraged to engage with the literature and help shaping the methodological choices. One direction would consist in a climate-informed contrastive approach, using meteorological state (e.g., accumulated temperature, day length, recent precipitation) as a proxy for phenological stage to construct training pairs across sites and years. Alternatively, JEPA-style predictive architectures training the model to anticipate future canopy states from current context would directly align pre-training with the early prediction task.

Goal

Contact Details

More information

Open this project...

Published since: 2026-04-30 , Earliest start: 2026-06-01 , Latest end: 2027-06-01

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Schindler Konrad

Topics Information, Computing and Communication Sciences , Biology

Implicit Earth: Neural Compression of Geospatial Foundation Embeddings

Photogrammetry and Remote Sensing (Prof. Schindler)

Global geospatial embedding products provide powerful, pre-computed latent representations of the Earth, but their discrete, grid-based nature requires significant storage and limits sub-pixel analysis. This thesis proposes training an Implicit Neural Representation (INR) to "memorize" these pre-computed embeddings. By training a coordinate-based network to map spatial coordinates to high-dimensional semantic vectors, this project investigates whether the entire latent representation of a region, country, or continent can be compressed into the network's weights while preserving the high-frequency semantics necessary for downstream tasks.

Keywords

Geospatial embeddings, implicit neural representation, location encoder, compression, foundation models.

Labels

Semester Project , Master Thesis

Description

Goal

Contact Details

More information

Open this project...

Published since: 2026-04-25 , Earliest start: 2026-06-01 , Latest end: 2027-04-30

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Scheibenreif Linus

Topics Information, Computing and Communication Sciences , Engineering and Technology

Synthesizing Aerial Images from Building Roof Geometry

Photogrammetry and Remote Sensing (Prof. Schindler)

This thesis explores a generative pipeline for creating accurate synthetic aerial imagery using existing geospatial data such as cadaster maps and OpenStreetMap (OSM). The proposed method investigates conditioning-based generative approaches to adapt building roof structures so they align with available labels while preserving the original aerial image style, including roof colors, textures, lighting and surroundings. The motivation is that many existing datasets contain misalignments and outdated annotations when compared to real imagery. By generating corrected yet realistic training data, this work aims to support future tasks such as roof segmentation, vectorization, and automated roof modeling.

Keywords

Synthetic Data Generation Remote Sensing Diffusion Models Geospatial Data Deep Learning Urban Mapping

Labels

Semester Project , Master Thesis

Description

Obtaining high-quality training data for Deep Learning models in remote sensing remains a significant challenge. Existing datasets, including cadaster data, OSM, and official government records, often exhibit discrepancies when compared to real-world aerial or satellite imagery. These misalignments arise due to differences in data acquisition methods, temporal inconsistencies, and inaccuracies in manually or semi-automatically generated vector datasets. This thesis investigates different conditioning strategies for diffusion-based generative models using structured inputs such as roof masks, polygons, or vector representations derived from geospatial data. In particular, spatial conditioning methods such as ControlNet [1], parameter- efficient adaptation techniques such as LoRA [2], and prompt-based conditioning approaches are considered as potential solutions. The focus is on maintaining high visual realism while enforcing strong geometric consistency between generated imagery and input roof structures. The resulting synthetic data is intended to support the development of more accurate methods for roof segmentation, roof vectorization, and automated roof modeling, while reducing the need for costly manual data correction. [1] Zhang, L., Rao, A. & Agrawala, M. (2023). Adding conditional control to text-to-image diffusion models. ICCV. [2] Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2022). Lora: Low- rank adaptation of large language models. ICLR. [3] Rombach, R., Blattmann, A., Lorenz, D., Esser, P. & Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. CVPR. **Requirements** - Background in data science and machine learning; - Knowledge of deep machine learning; - Programming in Python/PyTorch; - Experience with contemporary network architectures (vision transformers, diffusion models) is a plus **Location** On-site at the Photogrammetry and Remote Sensing Laboratory (Prof. Schindler, ETH Zürich)

Goal

Contact Details

More information

Open this project...

Published since: 2026-04-23 , Earliest start: 2026-09-01 , Latest end: 2027-08-31

Organization Photogrammetry and Remote Sensing (Prof. Schindler)

Hosts Schindler Konrad , Nicolicioiu Armand

Topics Information, Computing and Communication Sciences , Engineering and Technology