Konrad Schindler

Nando Metzger, Rodrigo Caye Daudt, Konrad Schindler

Performing super-resolution of a depth image using the guidance from an RGB image is a problem that concerns several fields, such as robotics, medical imaging, and remote sensing. While deep learning methods have achieved good results in this problem, recent work highlighted the value of combining modern methods with more formal frameworks. In this work, we propose a novel approach which combines guided anisotropic diffusion with a deep convolutional network and advances the state of the art for guided depth super-resolution. The edge transferring/enhancing properties of the diffusion are boosted by the contextual reasoning capabilities of modern networks, and a strict adjustment step guarantees perfect adherence to the source image. We achieve unprecedented results in three commonly used benchmarks for guided depth super-resolution. The performance gain compared to other methods is the largest at larger scales, such as x32 scaling. Code (https://github.com/prs-eth/Diffusion-Super-Resolution) for the proposed method is available to promote reproducibility of our results.

Yuanwen Yue, Theodora Kontogianni, Konrad Schindler et al.

We address 2D floorplan reconstruction from 3D scans. Existing approaches typically employ heuristically designed multi-stage pipelines. Instead, we formulate floorplan reconstruction as a single-stage structured prediction task: find a variable-size set of polygons, which in turn are variable-length sequences of ordered vertices. To solve it we develop a novel Transformer architecture that generates polygons of multiple rooms in parallel, in a holistic manner without hand-crafted intermediate stages. The model features two-level queries for polygons and corners, and includes polygon matching to make the network end-to-end trainable. Our method achieves a new state-of-the-art for two challenging datasets, Structured3D and SceneCAD, along with significantly faster inference than previous methods. Moreover, it can readily be extended to predict additional information, i.e., semantic room types and architectural elements like doors and windows. Our code and models are available at: https://github.com/ywyue/RoomFormer.

Liyuan Zhu, Yue Li, Erik Sandström et al. · stanford

Simultaneous Localization and Mapping (SLAM) based on 3D Gaussian Splats (3DGS) has recently shown promise towards more accurate, dense 3D scene maps. However, existing 3DGS-based methods fail to address the global consistency of the scene via loop closure and/or global bundle adjustment. To this end, we propose LoopSplat, which takes RGB-D images as input and performs dense mapping with 3DGS submaps and frame-to-model tracking. LoopSplat triggers loop closure online and computes relative loop edge constraints between submaps directly via 3DGS registration, leading to improvements in efficiency and accuracy over traditional global-to-local point cloud registration. It uses a robust pose graph optimization formulation and rigidly aligns the submaps to achieve global consistency. Evaluation on the synthetic Replica and real-world TUM-RGBD, ScanNet, and ScanNet++ datasets demonstrates competitive or superior tracking, mapping, and rendering compared to existing methods for dense RGB-D SLAM. Code is available at loopsplat.github.io.

Tao Sun, Yan Hao, Shengyu Huang et al. · eth-zurich, stanford

Building 3D geometric maps of man-made spaces is a well-established and active field that is fundamental to computer vision and robotics. However, considering the evolving nature of built environments, it is essential to question the capabilities of current mapping efforts in handling temporal changes. In addition, spatiotemporal mapping holds significant potential for achieving sustainability and circularity goals. Existing mapping approaches focus on small changes, such as object relocation or self-driving car operation; in all cases where the main structure of the scene remains fixed. Consequently, these approaches fail to address more radical changes in the structure of the built environment, such as geometry and topology. To this end, we introduce the Nothing Stands Still (NSS) benchmark, which focuses on the spatiotemporal registration of 3D scenes undergoing large spatial and temporal change, ultimately creating one coherent spatiotemporal map. Specifically, the benchmark involves registering two or more partial 3D point clouds (fragments) from the same scene but captured from different spatiotemporal views. In addition to the standard pairwise registration, we assess the multi-way registration of multiple fragments that belong to any temporal stage. As part of NSS, we introduce a dataset of 3D point clouds recurrently captured in large-scale building indoor environments that are under construction or renovation. The NSS benchmark presents three scenarios of increasing difficulty, to quantify the generalization ability of point cloud registration methods over space (within one building and across buildings) and time. We conduct extensive evaluations of state-of-the-art methods on NSS. The results demonstrate the necessity for novel methods specifically designed to handle large spatiotemporal changes. The homepage of our benchmark is at http://nothing-stands-still.com.

Nico Lang, Walter Jetz, Konrad Schindler et al.

The worldwide variation in vegetation height is fundamental to the global carbon cycle and central to the functioning of ecosystems and their biodiversity. Geospatially explicit and, ideally, highly resolved information is required to manage terrestrial ecosystems, mitigate climate change, and prevent biodiversity loss. Here, we present the first global, wall-to-wall canopy height map at 10 m ground sampling distance for the year 2020. No single data source meets these requirements: dedicated space missions like GEDI deliver sparse height data, with unprecedented coverage, whereas optical satellite images like Sentinel-2 offer dense observations globally, but cannot directly measure vertical structures. By fusing GEDI with Sentinel-2, we have developed a probabilistic deep learning model to retrieve canopy height from Sentinel-2 images anywhere on Earth, and to quantify the uncertainty in these estimates. The presented approach reduces the saturation effect commonly encountered when estimating canopy height from satellite images, allowing to resolve tall canopies with likely high carbon stocks. According to our map, only 5% of the global landmass is covered by trees taller than 30 m. Such data play an important role for conservation, e.g., we find that only 34% of these tall canopies are located within protected areas. Our model enables consistent, uncertainty-informed worldwide mapping and supports an ongoing monitoring to detect change and inform decision making. The approach can serve ongoing efforts in forest conservation, and has the potential to foster advances in climate, carbon, and biodiversity modelling.

Liyuan Zhu, Yuru Jia, Shengyu Huang et al. · stanford

Existing work on scene flow estimation focuses on autonomous driving and mobile robotics, while automated solutions are lacking for motion in nature, such as that exhibited by debris flows. We propose DEFLOW, a model for 3D motion estimation of debris flows, together with a newly captured dataset. We adopt a novel multi-level sensor fusion architecture and self-supervision to incorporate the inductive biases of the scene. We further adopt a multi-frame temporal processing module to enable flow speed estimation over time. Our model achieves state-of-the-art optical flow and depth estimation on our dataset, and fully automates the motion estimation for debris flows. The source code and dataset are available at project page.

Nikolai Kalischek, Rodrigo C. Daudt, Torben Peters et al.

The well-documented presence of texture bias in modern convolutional neural networks has led to a plethora of algorithms that promote an emphasis on shape cues, often to support generalization to new domains. Yet, common datasets, benchmarks and general model selection strategies are missing, and there is no agreed, rigorous evaluation protocol. In this paper, we investigate difficulties and limitations when training networks with reduced texture bias. In particular, we also show that proper evaluation and meaningful comparisons between methods are not trivial. We introduce BiasBed, a testbed for texture- and style-biased training, including multiple datasets and a range of existing algorithms. It comes with an extensive evaluation protocol that includes rigorous hypothesis testing to gauge the significance of the results, despite the considerable training instability of some style bias methods. Our extensive experiments, shed new light on the need for careful, statistically founded evaluation protocols for style bias (and beyond). E.g., we find that some algorithms proposed in the literature do not significantly mitigate the impact of style bias at all. With the release of BiasBed, we hope to foster a common understanding of consistent and meaningful comparisons, and consequently faster progress towards learning methods free of texture bias. Code is available at https://github.com/D1noFuzi/BiasBed

Nadine Rueegg, Silvia Zuffi, Konrad Schindler et al.

Our goal is to recover the 3D shape and pose of dogs from a single image. This is a challenging task because dogs exhibit a wide range of shapes and appearances, and are highly articulated. Recent work has proposed to directly regress the SMAL animal model, with additional limb scale parameters, from images. Our method, called BARC (Breed-Augmented Regression using Classification), goes beyond prior work in several important ways. First, we modify the SMAL shape space to be more appropriate for representing dog shape. But, even with a better shape model, the problem of regressing dog shape from an image is still challenging because we lack paired images with 3D ground truth. To compensate for the lack of paired data, we formulate novel losses that exploit information about dog breeds. In particular, we exploit the fact that dogs of the same breed have similar body shapes. We formulate a novel breed similarity loss consisting of two parts: One term encourages the shape of dogs from the same breed to be more similar than dogs of different breeds. The second one, a breed classification loss, helps to produce recognizable breed-specific shapes. Through ablation studies, we find that our breed losses significantly improve shape accuracy over a baseline without them. We also compare BARC qualitatively to WLDO with a perceptual study and find that our approach produces dogs that are significantly more realistic. This work shows that a-priori information about genetic similarity can help to compensate for the lack of 3D training data. This concept may be applicable to other animal species or groups of species. Our code is publicly available for research purposes at https://barc.is.tue.mpg.de/.

Sidi Wu, Yizi Chen, Konrad Schindler et al.

Historical maps provide useful spatio-temporal information on the Earth's surface before modern earth observation techniques came into being. To extract information from maps, neural networks, which gain wide popularity in recent years, have replaced hand-crafted map processing methods and tedious manual labor. However, aleatoric uncertainty, known as data-dependent uncertainty, inherent in the drawing/scanning/fading defects of the original map sheets and inadequate contexts when cropping maps into small tiles considering the memory limits of the training process, challenges the model to make correct predictions. As aleatoric uncertainty cannot be reduced even with more training data collected, we argue that complementary spatio-temporal contexts can be helpful. To achieve this, we propose a U-Net-based network that fuses spatio-temporal features with cross-attention transformers (U-SpaTem), aggregating information at a larger spatial range as well as through a temporal sequence of images. Our model achieves a better performance than other state-or-art models that use either temporal or spatial contexts. Compared with pure vision transformers, our model is more lightweight and effective. To the best of our knowledge, leveraging both spatial and temporal contexts have been rarely explored before in the segmentation task. Even though our application is on segmenting historical maps, we believe that the method can be transferred into other fields with similar problems like temporal sequences of satellite images. Our code is freely accessible at https://github.com/chenyizi086/wu.2023.sigspatial.git.

Shengyu Huang, Zan Gojcic, Jiahui Huang et al.

Multi-beam LiDAR sensors, as used on autonomous vehicles and mobile robots, acquire sequences of 3D range scans ("frames"). Each frame covers the scene sparsely, due to limited angular scanning resolution and occlusion. The sparsity restricts the performance of downstream processes like semantic segmentation or surface reconstruction. Luckily, when the sensor moves, frames are captured from a sequence of different viewpoints. This provides complementary information and, when accumulated in a common scene coordinate frame, yields a denser sampling and a more complete coverage of the underlying 3D scene. However, often the scanned scenes contain moving objects. Points on those objects are not correctly aligned by just undoing the scanner's ego-motion. In the present paper, we explore multi-frame point cloud accumulation as a mid-level representation of 3D scan sequences, and develop a method that exploits inductive biases of outdoor street scenes, including their geometric layout and object-level rigidity. Compared to state-of-the-art scene flow estimators, our proposed approach aims to align all 3D points in a common reference frame correctly accumulating the points on the individual objects. Our approach greatly reduces the alignment errors on several benchmark datasets. Moreover, the accumulated point clouds benefit high-level tasks like surface reconstruction.

Riccardo de Lutio, Alexander Becker, Stefano D'Aronco et al.

We introduce a novel formulation for guided super-resolution. Its core is a differentiable optimisation layer that operates on a learned affinity graph. The learned graph potentials make it possible to leverage rich contextual information from the guide image, while the explicit graph optimisation within the architecture guarantees rigorous fidelity of the high-resolution target to the low-resolution source. With the decision to employ the source as a constraint rather than only as an input to the prediction, our method differs from state-of-the-art deep architectures for guided super-resolution, which produce targets that, when downsampled, will only approximately reproduce the source. This is not only theoretically appealing, but also produces crisper, more natural-looking images. A key property of our method is that, although the graph connectivity is restricted to the pixel lattice, the associated edge potentials are learned with a deep feature extractor and can encode rich context information over large receptive fields. By taking advantage of the sparse graph connectivity, it becomes possible to propagate gradients through the optimisation layer and learn the edge potentials from data. We extensively evaluate our method on several datasets, and consistently outperform recent baselines in terms of quantitative reconstruction errors, while also delivering visually sharper outputs. Moreover, we demonstrate that our method generalises particularly well to new datasets not seen during training.

Andrés C. Rodríguez, Stefano D'Aronco, Rodrigo Caye Daudt et al.

We exploit field guides to learn bird species recognition, in particular zero-shot recognition of unseen species. Illustrations contained in field guides deliberately focus on discriminative properties of each species, and can serve as side information to transfer knowledge from seen to unseen bird species. We study two approaches: (1) a contrastive encoding of illustrations, which can be fed into standard zero-shot learning schemes; and (2) a novel method that leverages the fact that illustrations are also images and as such structurally more similar to photographs than other kinds of side information. Our results show that illustrations from field guides, which are readily available for a wide range of species, are indeed a competitive source of side information for zero-shot learning. On a subset of the iNaturalist2021 dataset with 749 seen and 739 unseen species, we obtain a classification accuracy of unseen bird species of $12\%$ @top-1 and $38\%$ @top-10, which shows the potential of field guides for challenging real-world scenarios with many species. Our code is available at https://github.com/ac-rodriguez/zsl_billow

Ye Hong, Yatao Zhang, Konrad Schindler et al.

Accurate activity location prediction is a crucial component of many mobility applications and is particularly required to develop personalized, sustainable transportation systems. Despite the widespread adoption of deep learning models, next location prediction models lack a comprehensive discussion and integration of mobility-related spatio-temporal contexts. Here, we utilize a multi-head self-attentional (MHSA) neural network that learns location transition patterns from historical location visits, their visit time and activity duration, as well as their surrounding land use functions, to infer an individual's next location. Specifically, we adopt point-of-interest data and latent Dirichlet allocation for representing locations' land use contexts at multiple spatial scales, generate embedding vectors of the spatio-temporal features, and learn to predict the next location with an MHSA network. Through experiments on two large-scale GNSS tracking datasets, we demonstrate that the proposed model outperforms other state-of-the-art prediction models, and reveal the contribution of various spatio-temporal contexts to the model's performance. Moreover, we find that the model trained on population data achieves higher prediction performance with fewer parameters than individual-level models due to learning from collective movement patterns. We also reveal mobility conducted in the recent past and one week before has the largest influence on the current prediction, showing that learning from a subset of the historical mobility is sufficient to obtain an accurate location prediction result. We believe that the proposed model is vital for context-aware mobility prediction. The gained insights will help to understand location prediction models and promote their implementation for mobility applications.

Mehmet Ozgur Turkoglu, Alexander Becker, Hüseyin Anil Gündüz et al.

The ability to estimate epistemic uncertainty is often crucial when deploying machine learning in the real world, but modern methods often produce overconfident, uncalibrated uncertainty predictions. A common approach to quantify epistemic uncertainty, usable across a wide class of prediction models, is to train a model ensemble. In a naive implementation, the ensemble approach has high computational cost and high memory demand. This challenges in particular modern deep learning, where even a single deep network is already demanding in terms of compute and memory, and has given rise to a number of attempts to emulate the model ensemble without actually instantiating separate ensemble members. We introduce FiLM-Ensemble, a deep, implicit ensemble method based on the concept of Feature-wise Linear Modulation (FiLM). That technique was originally developed for multi-task learning, with the aim of decoupling different tasks. We show that the idea can be extended to uncertainty quantification: by modulating the network activations of a single deep network with FiLM, one obtains a model ensemble with high diversity, and consequently well-calibrated estimates of epistemic uncertainty, with low computational overhead in comparison. Empirically, FiLM-Ensemble outperforms other implicit ensemble methods, and it and comes very close to the upper bound of an explicit ensemble of networks (sometimes even beating it), at a fraction of the memory cost.

Yuanwen Yue, Sabarinath Mahadevan, Jonas Schult et al.

During interactive segmentation, a model and a user work together to delineate objects of interest in a 3D point cloud. In an iterative process, the model assigns each data point to an object (or the background), while the user corrects errors in the resulting segmentation and feeds them back into the model. The current best practice formulates the problem as binary classification and segments objects one at a time. The model expects the user to provide positive clicks to indicate regions wrongly assigned to the background and negative clicks on regions wrongly assigned to the object. Sequentially visiting objects is wasteful since it disregards synergies between objects: a positive click for a given object can, by definition, serve as a negative click for nearby objects. Moreover, a direct competition between adjacent objects can speed up the identification of their common boundary. We introduce AGILE3D, an efficient, attention-based model that (1) supports simultaneous segmentation of multiple 3D objects, (2) yields more accurate segmentation masks with fewer user clicks, and (3) offers faster inference. Our core idea is to encode user clicks as spatial-temporal queries and enable explicit interactions between click queries as well as between them and the 3D scene through a click attention module. Every time new clicks are added, we only need to run a lightweight decoder that produces updated segmentation masks. In experiments with four different 3D point cloud datasets, AGILE3D sets a new state-of-the-art. Moreover, we also verify its practicality in real-world setups with real user studies.

Yuanwen Yue, Damien Robert, Jianyuan Wang et al.

Modern neural architectures for 3D point cloud processing contain both convolutional layers and attention blocks, but the best way to assemble them remains unclear. We analyse the role of different computational blocks in 3D point cloud networks and find an intuitive behaviour: convolution is adequate to extract low-level geometry at high-resolution in early layers, where attention is expensive without bringing any benefits; attention captures high-level semantics and context in low-resolution, deep layers more efficiently. Guided by this design principle, we propose a new, improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers. To avoid the loss of spatial layout information when discarding redundant convolution layers, we introduce a novel, training-free 3D positional encoding, PointROPE. The resulting LitePT model has $3.6\times$ fewer parameters, runs $2\times$ faster, and uses $2\times$ less memory than the state-of-the-art Point Transformer V3, but nonetheless matches or even outperforms it on a range of tasks and datasets. Code and models are available at: https://github.com/prs-eth/LitePT.

Tianfu Wang, Menelaos Kanakis, Konrad Schindler et al.

Diffusion-based text-to-image models ignited immense attention from the vision community, artists, and content creators. Broad adoption of these models is due to significant improvement in the quality of generations and efficient conditioning on various modalities, not just text. However, lifting the rich generative priors of these 2D models into 3D is challenging. Recent works have proposed various pipelines powered by the entanglement of diffusion models and neural fields. We explore the power of pretrained 2D diffusion models and standard 3D neural radiance fields as independent, standalone tools and demonstrate their ability to work together in a non-learned fashion. Such modularity has the intrinsic advantage of eased partial upgrades, which became an important property in such a fast-paced domain. Our pipeline accepts any legacy renderable geometry, such as textured or untextured meshes, orchestrates the interaction between 2D generative refinement and 3D consistency enforcement tools, and outputs a painted input geometry in several formats. We conduct a large-scale study on a wide range of objects and categories from the ShapeNetSem dataset and demonstrate the advantages of our approach, both qualitatively and quantitatively. Project page: https://www.obukhov.ai/repainting_3d_assets

Theodora Kontogianni, Ekin Celikkan, Siyu Tang et al.

We propose an interactive approach for 3D instance segmentation, where users can iteratively collaborate with a deep learning model to segment objects in a 3D point cloud directly. Current methods for 3D instance segmentation are generally trained in a fully-supervised fashion, which requires large amounts of costly training labels, and does not generalize well to classes unseen during training. Few works have attempted to obtain 3D segmentation masks using human interactions. Existing methods rely on user feedback in the 2D image domain. As a consequence, users are required to constantly switch between 2D images and 3D representations, and custom architectures are employed to combine multiple input modalities. Therefore, integration with existing standard 3D models is not straightforward. The core idea of this work is to enable users to interact directly with 3D point clouds by clicking on desired 3D objects of interest~(or their background) to interactively segment the scene in an open-world setting. Specifically, our method does not require training data from any target domain, and can adapt to new environments where no appropriate training sets are available. Our system continuously adjusts the object segmentation based on the user feedback and achieves accurate dense 3D segmentation masks with minimal human effort (few clicks per object). Besides its potential for efficient labeling of large-scale and varied 3D datasets, our approach, where the user directly interacts with the 3D environment, enables new applications in AR/VR and human-robot interaction.

Mikhail Usvyatsov, Rafael Ballester-Ripoll, Konrad Schindler

We present tntorch, a tensor learning framework that supports multiple decompositions (including Candecomp/Parafac, Tucker, and Tensor Train) under a unified interface. With our library, the user can learn and handle low-rank tensors with automatic differentiation, seamless GPU support, and the convenience of PyTorch's API. Besides decomposition algorithms, tntorch implements differentiable tensor algebra, rank truncation, cross-approximation, batch processing, comprehensive tensor arithmetics, and more.

Nando Metzger, John E. Vargas-Muñoz, Rodrigo C. Daudt et al.

Fine-grained population maps are needed in several domains, like urban planning, environmental monitoring, public health, and humanitarian operations. Unfortunately, in many countries only aggregate census counts over large spatial units are collected, moreover, these are not always up-to-date. We present POMELO, a deep learning model that employs coarse census counts and open geodata to estimate fine-grained population maps with 100m ground sampling distance. Moreover, the model can also estimate population numbers when no census counts at all are available, by generalizing across countries. In a series of experiments for several countries in sub-Saharan Africa, the maps produced with POMELOare in good agreement with the most detailed available reference counts: disaggregation of coarse census counts reaches R2 values of 85-89%; unconstrained prediction in the absence of any counts reaches 48-69%.

Binbin Xiang, Torben Peters, Theodora Kontogianni et al.

Panoptic segmentation is the combination of semantic and instance segmentation: assign the points in a 3D point cloud to semantic categories and partition them into distinct object instances. It has many obvious applications for outdoor scene understanding, from city mapping to forest management. Existing methods struggle to segment nearby instances of the same semantic category, like adjacent pieces of street furniture or neighbouring trees, which limits their usability for inventory- or management-type applications that rely on object instances. This study explores the steps of the panoptic segmentation pipeline concerned with clustering points into object instances, with the goal to alleviate that bottleneck. We find that a carefully designed clustering strategy, which leverages multiple types of learned point embeddings, significantly improves instance segmentation. Experiments on the NPM3D urban mobile mapping dataset and the FOR-instance forest dataset demonstrate the effectiveness and versatility of the proposed strategy.

Binbin Xiang, Yuanwen Yue, Torben Peters et al.

3D point cloud panoptic segmentation is the combined task to (i) assign each point to a semantic class and (ii) separate the points in each class into object instances. Recently there has been an increased interest in such comprehensive 3D scene understanding, building on the rapid advances of semantic segmentation due to the advent of deep 3D neural networks. Yet, to date there is very little work about panoptic segmentation of outdoor mobile-mapping data, and no systematic comparisons. The present paper tries to close that gap. It reviews the building blocks needed to assemble a panoptic segmentation pipeline and the related literature. Moreover, a modular pipeline is set up to perform comprehensive, systematic experiments to assess the state of panoptic segmentation in the context of street mapping. As a byproduct, we also provide the first public dataset for that task, by extending the NPM3D dataset to include instance labels. That dataset and our source code are publicly available. We discuss which adaptations are need to adapt current panoptic segmentation methods to outdoor scenes and large objects. Our study finds that for mobile mapping data, KPConv performs best but is slower, while PointNet++ is fastest but performs significantly worse. Sparse CNNs are in between. Regardless of the backbone, Instance segmentation by clustering embedding features is better than using shifted coordinates.

Nikolai Kalischek, Torben Peters, Jan D. Wegner et al.

Probabilistic denoising diffusion models (DDMs) have set a new standard for 2D image generation. Extending DDMs for 3D content creation is an active field of research. Here, we propose TetraDiffusion, a diffusion model that operates on a tetrahedral partitioning of 3D space to enable efficient, high-resolution 3D shape generation. Our model introduces operators for convolution and transpose convolution that act directly on the tetrahedral partition, and seamlessly includes additional attributes such as color. Remarkably, TetraDiffusion enables rapid sampling of detailed 3D objects in nearly real-time with unprecedented resolution. It's also adaptable for generating 3D shapes conditioned on 2D images. Compared to existing 3D mesh diffusion techniques, our method is up to 200 times faster in inference speed, works on standard consumer hardware, and delivers superior results.

Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis et al.

Learning neural fields has been an active topic in deep learning research, focusing, among other issues, on finding more compact and easy-to-fit representations. In this paper, we introduce a novel low-rank representation termed Tensor Train Neural Fields (TT-NF) for learning neural fields on dense regular grids and efficient methods for sampling from them. Our representation is a TT parameterization of the neural field, trained with backpropagation to minimize a non-convex objective. We analyze the effect of low-rank compression on the downstream task quality metrics in two settings. First, we demonstrate the efficiency of our method in a sandbox task of tensor denoising, which admits comparison with SVD-based schemes designed to minimize reconstruction error. Furthermore, we apply the proposed approach to Neural Radiance Fields, where the low-rank structure of the field corresponding to the best quality can be discovered only through learning.

Nando Metzger, Rodrigo Caye Daudt, Devis Tuia et al.

Detailed population maps play an important role in diverse fields ranging from humanitarian action to urban planning. Generating such maps in a timely and scalable manner presents a challenge, especially in data-scarce regions. To address it we have developed POPCORN, a population mapping method whose only inputs are free, globally available satellite images from Sentinel-1 and Sentinel-2; and a small number of aggregate population counts over coarse census districts for calibration. Despite the minimal data requirements our approach surpasses the mapping accuracy of existing schemes, including several that rely on building footprints derived from high-resolution imagery. E.g., we were able to produce population maps for Rwanda with 100m GSD based on less than 400 regional census counts. In Kigali, those maps reach an R^2 score of 66% w.r.t. a ground truth reference map, with an average error of only about 10 inhabitants/ha. Conveniently, POPCORN retrieves explicit maps of built-up areas and of local building occupancy rates, making the mapping process interpretable and offering additional insights, for instance about the distribution of built-up, but unpopulated areas, e.g., industrial warehouses. Moreover, we find that, once trained, the model can be applied repeatedly to track population changes; and that it can be transferred to geographically similar regions, e.g., from Uganda to Rwanda). With our work we aim to democratize access to up-to-date and high-resolution population maps, recognizing that some regions faced with particularly strong population dynamics may lack the resources for costly micro-census campaigns.

Nikolai Kalischek, Nico Lang, Cécile Renier et al.

Côte d'Ivoire and Ghana, the world's largest producers of cocoa, account for two thirds of the global cocoa production. In both countries, cocoa is the primary perennial crop, providing income to almost two million farmers. Yet precise maps of cocoa planted area are missing, hindering accurate quantification of expansion in protected areas, production and yields, and limiting information available for improved sustainability governance. Here, we combine cocoa plantation data with publicly available satellite imagery in a deep learning framework and create high-resolution maps of cocoa plantations for both countries, validated in situ. Our results suggest that cocoa cultivation is an underlying driver of over 37% and 13% of forest loss in protected areas in Côte d'Ivoire and Ghana, respectively, and that official reports substantially underestimate the planted area, up to 40% in Ghana. These maps serve as a crucial building block to advance understanding of conservation and economic development in cocoa producing regions.

Mikhail Usvyatsov, Rafael Ballester-Rippoll, Lina Bashaeva et al.

Unlike 2D raster images, there is no single dominant representation for 3D visual data processing. Different formats like point clouds, meshes, or implicit functions each have their strengths and weaknesses. Still, grid representations such as signed distance functions have attractive properties also in 3D. In particular, they offer constant-time random access and are eminently suitable for modern machine learning. Unfortunately, the storage size of a grid grows exponentially with its dimension. Hence they often exceed memory limits even at moderate resolution. This work proposes using low-rank tensor formats, including the Tucker, tensor train, and quantics tensor train decompositions, to compress time-varying 3D data. Our method iteratively computes, voxelizes, and compresses each frame's truncated signed distance function and applies tensor rank truncation to condense all frames into a single, compressed tensor that represents the entire 4D scene. We show that low-rank tensor compression is extremely compact to store and query time-varying signed distance functions. It significantly reduces the memory footprint of 4D scenes while remarkably preserving their geometric quality. Unlike existing, iterative learning-based approaches like DeepSDF and NeRF, our method uses a closed-form algorithm with theoretical guarantees.

Han Sun, Rui Gong, Konrad Schindler et al.

Domain adaptive object detection aims to leverage the knowledge learned from a labeled source domain to improve the performance on an unlabeled target domain. Prior works typically require the access to the source domain data for adaptation, and the availability of sufficient data on the target domain. However, these assumptions may not hold due to data privacy and rare data collection. In this paper, we propose and investigate a more practical and challenging domain adaptive object detection problem under both source-free and few-shot conditions, named as SF-FSDA. To overcome this problem, we develop an efficient labeled data factory based approach. Without accessing the source domain, the data factory renders i) infinite amount of synthesized target-domain like images, under the guidance of the few-shot image samples and text description from the target domain; ii) corresponding bounding box and category annotations, only demanding minimum human effort, i.e., a few manually labeled examples. On the one hand, the synthesized images mitigate the knowledge insufficiency brought by the few-shot condition. On the other hand, compared to the popular pseudo-label technique, the generated annotations from data factory not only get rid of the reliance on the source pretrained object detection model, but also alleviate the unavoidably pseudo-label noise due to domain shift and source-free condition. The generated dataset is further utilized to adapt the source pretrained object detection model, realizing the robust object detection under SF-FSDA. The experiments on different settings showcase that our proposed approach outperforms other state-of-the-art methods on SF-FSDA problem. Our codes and models will be made publicly available.

Alexander Becker, Rodrigo Caye Daudt, Dominik Narnhofer et al.

Recent approaches to arbitrary-scale single image super-resolution (ASR) use neural fields to represent continuous signals that can be sampled at arbitrary resolutions. However, point-wise queries of neural fields do not naturally match the point spread function (PSF) of pixels, which may cause aliasing in the super-resolved image. Existing methods attempt to mitigate this by approximating an integral version of the field at each scaling factor, compromising both fidelity and generalization. In this work, we introduce neural heat fields, a novel neural field formulation that inherently models a physically exact PSF. Our formulation enables analytically correct anti-aliasing at any desired output resolution, and -- unlike supersampling -- at no additional cost. Building on this foundation, we propose Thera, an end-to-end ASR method that substantially outperforms existing approaches, while being more parameter-efficient and offering strong theoretical guarantees. The project page is at https://therasr.github.io.

Yuchang Jiang, Marius Rüetschi, Vivien Sainte Fare Garnot et al.

Monitoring and understanding forest dynamics is essential for environmental conservation and management. This is why the Swiss National Forest Inventory (NFI) provides countrywide vegetation height maps at a spatial resolution of 0.5 m. Its long update time of 6 years, however, limits the temporal analysis of forest dynamics. This can be improved by using spaceborne remote sensing and deep learning to generate large-scale vegetation height maps in a cost-effective way. In this paper, we present an in-depth analysis of these methods for operational application in Switzerland. We generate annual, countrywide vegetation height maps at a 10-meter ground sampling distance for the years 2017 to 2020 based on Sentinel-2 satellite imagery. In comparison to previous works, we conduct a large-scale and detailed stratified analysis against a precise Airborne Laser Scanning reference dataset. This stratified analysis reveals a close relationship between the model accuracy and the topology, especially slope and aspect. We assess the potential of deep learning-derived height maps for change detection and find that these maps can indicate changes as small as 250 $m^2$. Larger-scale changes caused by a winter storm are detected with an F1-score of 0.77. Our results demonstrate that vegetation height maps computed from satellite imagery with deep learning are a valuable, complementary, cost-effective source of evidence to increase the temporal resolution for national forest assessments.

Xiang Zhang, Bingxin Ke, Hayko Riemenschneider et al.

By training over large-scale datasets, zero-shot monocular depth estimation (MDE) methods show robust performance in the wild but often suffer from insufficient detail. Although recent diffusion-based MDE approaches exhibit a superior ability to extract details, they struggle in geometrically complex scenes that challenge their geometry prior, trained on less diverse 3D data. To leverage the complementary merits of both worlds, we propose BetterDepth to achieve geometrically correct affine-invariant MDE while capturing fine details. Specifically, BetterDepth is a conditional diffusion-based refiner that takes the prediction from pre-trained MDE models as depth conditioning, in which the global depth layout is well-captured, and iteratively refines details based on the input image. For the training of such a refiner, we propose global pre-alignment and local patch masking methods to ensure BetterDepth remains faithful to the depth conditioning while learning to add fine-grained scene details. With efficient training on small-scale synthetic datasets, BetterDepth achieves state-of-the-art zero-shot MDE performance on diverse public datasets and on in-the-wild scenes. Moreover, BetterDepth can improve the performance of other MDE models in a plug-and-play manner without further re-training.

Niccolo Avogaro, Nayanika Debnath, Li Mi et al. · ibm-research

Despite recent successes, test-time scaling - i.e., dynamically expanding the token budget during inference as needed - remains brittle for vision-language models (VLMs): unstructured chains-of-thought about images entangle perception and reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers. Moreover, expensive reinforcement learning with hand-crafted rewards is required to achieve good performance. Here, we introduce SPARC (Separating Perception And Reasoning Circuits), a modular framework that explicitly decouples visual perception from reasoning. Inspired by sequential sensory-to-cognitive processing in the brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric compute allocation (e.g., prioritizing perceptual processing under distribution shift), supports selective optimization (e.g., improving the perceptual stage alone when it is the bottleneck for end-to-end performance), and accommodates compressed contexts by running global search at lower image resolutions and allocating high-resolution processing only to selected regions, thereby reducing total visual tokens count and compute. Across challenging visual reasoning benchmarks, SPARC outperforms monolithic baselines and strong visual-grounding approaches. For instance, SPARC improves the accuracy of Qwen3VL-4B on the $V^*$ VQA benchmark by 6.7 percentage points, and it surpasses "thinking with images" by 4.6 points on a challenging OOD task despite requiring a 200$\times$ lower token budget.

Bingxin Ke, Qunjie Zhou, Jiahui Huang et al. · nvidia, utoronto

We introduce CAPA, a parameter-efficient test-time optimization framework that adapts pre-trained 3D foundation models (FMs) for depth completion, using sparse geometric cues. Unlike prior methods that train task-specific encoders for auxiliary inputs, which often overfit and generalize poorly, CAPA freezes the FM backbone. Instead, it updates only a minimal set of parameters using Parameter-Efficient Fine-Tuning (e.g. LoRA or VPT), guided by gradients calculated directly from the sparse observations available at inference time. This approach effectively grounds the foundation model's geometric prior in the scene-specific measurements, correcting distortions and misplaced structures. For videos, CAPA introduces sequence-level parameter sharing, jointly adapting all frames to exploit temporal correlations, improve robustness, and enforce multi-frame consistency. CAPA is model-agnostic, compatible with any ViT-based FM, and achieves state-of-the-art results across diverse condition patterns on both indoor and outdoor datasets. Project page: research.nvidia.com/labs/dvl/projects/capa.

Vukasin Bozic, Isidora Slavkovic, Dominik Narnhofer et al.

Geometry estimation from perspective images has greatly advanced, maturing to the point where off-the-shelf foundation models are able to reconstruct 3D scene structure not only from multi-view imagery, but even from a single view. A natural extension is 3D reconstruction from panoramas, with the exciting prospect of recovering a full 360-degree scene from a single panoramic image. In this work, we introduce PaGeR (Panoramic Geometry Reconstruction), a framework to lift powerful 3D foundation models designed for perspective imagery to the panorama domain. Our strategy is to start from a pre-trained transformer for 3D reconstruction and turn it into a unified high-performance model that predicts scale-invariant depth, metric depth, surface normals, and sky masks from both perspective and omnidirectional images, in a single forward pass. By keeping architectural changes to a minimum and mixing perspective and panoramic images during training, PaGeR retains the rich 3D prior of the underlying foundation model while learning to also estimate geometrically consistent 360-degree scenes from single panoramas. We extensively test our method in both indoor and outdoor environments and find that it delivers state-of-the-art performance and excellent zero-shot performance across a wide range of scenes.

Yuchang Jiang, Anton Raichuk, Xiaoye Tong et al.

Monitoring tree crop expansion is vital for zero-deforestation policies like the European Union's Regulation on Deforestation-free Products (EUDR). However, these efforts are hindered by a lack of highresolution data distinguishing diverse agricultural systems from forests. Here, we present the first 10m-resolution tree crop map for South America, generated using a multi-modal, spatio-temporal deep learning model trained on Sentinel-1 and Sentinel-2 satellite imagery time series. The map identifies approximately 11 million hectares of tree crops, 23% of which is linked to 2000-2020 forest cover loss. Critically, our analysis reveals that existing regulatory maps supporting the EUDR often classify established agriculture, particularly smallholder agroforestry, as "forest". This discrepancy risks false deforestation alerts and unfair penalties for small-scale farmers. Our work mitigates this risk by providing a high-resolution baseline, supporting conservation policies that are effective, inclusive, and equitable.

Olivier Dietrich, Merlin Alfredsson, Emilia Arens et al.

Natural disasters demand rapid damage assessment to guide humanitarian response. Here, we investigate whether medium-resolution Earth observation images from the Copernicus program can support building damage assessment, complementing very-high resolution imagery with often limited availability. We introduce xBD-S12, a dataset of 10,315 pre- and post-disaster image pairs from both Sentinel-1 and Sentinel-2, spatially and temporally aligned with the established xBD benchmark. In a series of experiments, we demonstrate that building damage can be detected and mapped rather well in many disaster scenarios, despite the moderate 10$\,$m ground sampling distance. We also find that, for damage mapping at that resolution, architectural sophistication does not seem to bring much advantage: more complex model architectures tend to struggle with generalization to unseen disasters, and geospatial foundation models bring little practical benefit. Our results suggest that Copernicus images are a viable data source for rapid, wide-area damage assessment and could play an important role alongside VHR imagery. We release the xBD-S12 dataset, code, and trained models to support further research.

Tjark Behrens, Anton Obukhov, Bingxin Ke et al.

We introduce StereoSpace, a diffusion-based framework for monocular-to-stereo synthesis that models geometry purely through viewpoint conditioning, without explicit depth or warping. A canonical rectified space and the conditioning guide the generator to infer correspondences and fill disocclusions end-to-end. To ensure fair and leakage-free evaluation, we introduce an end-to-end protocol that excludes any ground truth or proxy geometry estimates at test time. The protocol emphasizes metrics reflecting downstream relevance: iSQoE for perceptual comfort and MEt3R for geometric consistency. StereoSpace surpasses other methods from the warp & inpaint, latent-warping, and warped-conditioning categories, achieving sharp parallax and strong robustness on layered and non-Lambertian scenes. This establishes viewpoint-conditioned diffusion as a scalable, depth-free solution for stereo generation.

Sidi Wu, Yizi Chen, Maurizio Gribaudi et al.

Prior to modern Earth observation technologies, historical maps provide a unique record of long-term urban transformation and offer a lens on the evolving identity of cities. However, extracting consistent and fine-grained change information from historical map series remains challenging due to spatial misalignment, cartographic variation, and degrading document quality, limiting most analyses to small-scale or qualitative approaches. We propose a fully automated, deep learning-based framework for fine-grained urban change analysis from large collections of historical maps, built on a modular design that integrates dense map alignment, multi-temporal object detection, and change profiling. This framework shifts the analysis of historical maps from ad hoc visual comparison toward systematic, quantitative characterization of urban change. Experiments demonstrate the robust performance of the proposed alignment and object detection methods. Applied to Paris between 1868 and 1937, the framework reveals the spatial and temporal heterogeneity in urban transformation, highlighting its relevance for research in the social sciences and humanities. The modular design of our framework further supports adaptation to diverse cartographic contexts and downstream applications.

Fabian Dionys Schrag, Mehmet Ozgur Turkoglu, Konrad Schindler et al.

Domain adaptation (DA) addresses the challenge of transferring a machine learning model trained on a source domain to a target domain with a different data distribution. In this work, we study DA for the task of Rumex obtusifolius (Rumex) image classification. We train models on a published, ground vehicle-based dataset (source) and evaluate their performance on a custom target dataset acquired by unmanned aerial vehicles (UAVs). We find that Convolutional Neural Network (CNN) models, specifically ResNets, generalize poorly to the target domain, even after fine-tuning on the source data. Applying moment-matching and maximum classifier discrepancy, two established DA techniques, substantially improves target-domain performance. However, Vision Transformer (ViT) models pretrained with self-supervised objectives (DINOv2, DINOv3) handle domain shifts intrinsically well, surpassing even moment-matching-trained ResNets, likely due to the rich, general-purpose representations acquired during large-scale pretraining. Using ViTs fine-tuned on the source dataset, we demonstrate high classification performances in the range of F1=0.8 on our target dataset. To support further research on DA for weed detection in grassland systems, we publicly release our UAV-based target dataset AGSMultiRumex, comprising data from 15 flights over Swiss meadows.

Yue Ma, Zhikai Wang, Tianhao Ren et al.

Video motion transfer aims to synthesize videos by generating visual content according to a text prompt while transferring the motion pattern observed in a reference video. Recent methods predominantly use the Diffusion Transformer (DiT) architecture. To achieve satisfactory runtime, several methods attempt to accelerate the computations in the DiT, but fail to address structural sources of inefficiency. In this work, we identify and remove two types of computational redundancy in earlier work: motion redundancy arises because the generic DiT architecture does not reflect the fact that frame-to-frame motion is small and smooth; gradient redundancy occurs if one ignores that gradients change slowly along the diffusion trajectory. To mitigate motion redundancy, we mask the corresponding attention layers to a local neighborhood such that interaction weights are not computed unnecessarily distant image regions. To exploit gradient redundancy, we design an optimization scheme that reuses gradients from previous diffusion steps and skips unwarranted gradient computations. On average, FastVMT achieves a 3.43x speedup without degrading the visual fidelity or the temporal consistency of the generated videos.

Hongruixuan Chen, Jian Song, Olivier Dietrich et al.

Disaster events occur around the world and cause significant damage to human life and property. Earth observation (EO) data enables rapid and comprehensive building damage assessment (BDA), an essential capability in the aftermath of a disaster to reduce human casualties and to inform disaster relief efforts. Recent research focuses on the development of AI models to achieve accurate mapping of unseen disaster events, mostly using optical EO data. However, solutions based on optical data are limited to clear skies and daylight hours, preventing a prompt response to disasters. Integrating multimodal (MM) EO data, particularly the combination of optical and SAR imagery, makes it possible to provide all-weather, day-and-night disaster responses. Despite this potential, the development of robust multimodal AI models has been constrained by the lack of suitable benchmark datasets. In this paper, we present a BDA dataset using veRy-hIGH-resoluTion optical and SAR imagery (BRIGHT) to support AI-based all-weather disaster response. To the best of our knowledge, BRIGHT is the first open-access, globally distributed, event-diverse MM dataset specifically curated to support AI-based disaster response. It covers five types of natural disasters and two types of man-made disasters across 14 regions worldwide, with a particular focus on developing countries where external assistance is most needed. The optical and SAR imagery in BRIGHT, with a spatial resolution between 0.3-1 meters, provides detailed representations of individual buildings, making it ideal for precise BDA. In our experiments, we have tested seven advanced AI models trained with our BRIGHT to validate the transferability and robustness. The dataset and code are available at https://github.com/ChenHongruixuan/BRIGHT. BRIGHT also serves as the official dataset for the 2025 IEEE GRSS Data Fusion Contest.

Nando Metzger, Mehmet Özgür Türkoglu, Rodrigo Caye Daudt et al.

Forecasting where and when new buildings will emerge is a rather unexplored topic, but one that is very useful in many disciplines such as urban planning, agriculture, resource management, and even autonomous flying. In the present work, we present a method that accomplishes this task with a deep neural network and a custom pretraining procedure. In Stage 1, a U-Net backbone is pretrained within a Siamese network architecture that aims to solve a (building) change detection task. In Stage 2, the backbone is repurposed to forecast the emergence of new buildings based solely on one image acquired before its construction. Furthermore, we also present a model that forecasts the time range within which the change will occur. We validate our approach using the SpaceNet7 dataset, which covers an area of 960 km^2 at 24 points in time across two years. In our experiments, we found that our proposed pretraining method consistently outperforms the traditional pretraining using the ImageNet dataset. We also show that it is to some degree possible to predict in advance when building changes will occur.

Sidi Wu, Yizi Chen, Samuel Mermet et al.

Most image-to-image translation models postulate that a unique correspondence exists between the semantic classes of the source and target domains. However, this assumption does not always hold in real-world scenarios due to divergent distributions, different class sets, and asymmetrical information representation. As conventional GANs attempt to generate images that match the distribution of the target domain, they may hallucinate spurious instances of classes absent from the source domain, thereby diminishing the usefulness and reliability of translated images. CycleGAN-based methods are also known to hide the mismatched information in the generated images to bypass cycle consistency objectives, a process known as steganography. In response to the challenge of non-bijective image translation, we introduce StegoGAN, a novel model that leverages steganography to prevent spurious features in generated images. Our approach enhances the semantic consistency of the translated images without requiring additional postprocessing or supervision. Our experimental evaluations demonstrate that StegoGAN outperforms existing GAN-based models across various non-bijective image-to-image translation tasks, both qualitatively and quantitatively. Our code and pretrained models are accessible at https://github.com/sian-wusidi/StegoGAN.

Hyojun Go, Hyungjin Chung, Prune Truong et al.

For practical use, diffusion- or flow-based generative models must be aligned with task-specific rewards, such as prompt fidelity or aesthetic preference. That alignment is challenging because the reward is defined for clean output images, but the alignment procedure requires value function estimates at noisy intermediate latents. Existing methods resort to Tweedie-style or Monte Carlo approximations, trading off estimator bias against computational cost: Tweedie estimates are efficient but biased, while Monte Carlo estimates are more accurate but require expensive rollouts. A natural alternative would be a learned value function, but it remains an open question how to effectively train a strong and general value model specifically for noisy latents. Here, we propose StitchVM, a model stitching framework that efficiently transfers reward models pretrained for clean images to the noisy latent regime. StitchVM starts from an existing, truncated pixel-space reward model and attaches a frozen diffusion backbone to it as its head. From the pixel-space model, the resulting hybrid retains a carefully pretrained, robust reward capability; from the diffusion backbone, it inherits its native ability to handle noisy latents. The stitching procedure is exceptionally lightweight, e.g., stitching and finetuning CLIP ViT-L and SD 3.5 Medium takes only 10 GPU-hours. By lifting powerful pixel-space reward models to latent space, StitchVM opens up a new style of diffusion alignment: instead of rough, yet costly per-sample approximation of the value function, the correct function for the actual, noisy latents is constructed once and then amortized over many samples and iterations. We show that this approach yields improvements across a broad range of downstream steering and post-training methods: DPS becomes $3.2\times$ faster while halving peak GPU memory, and DiffusionNFT becomes $2.3\times$ faster.

Nicola Farronato, Niccolo Avogaro, Thomas Frick et al.

Automated structural health monitoring is essential to prevent catastrophic infrastructure failures. Precise, pixel-level defect segmentation is needed to accurately assess structural integrity, but progress in defect segmentation for civil infrastructures has been held back by an extreme scarcity of data, which requires costly expert annotation. The need for data is accentuated by algorithmic hurdles intrinsic to the problem, including center-bias and the need to rely more on shape when inspecting nearly textureless building materials. To remove the bottleneck, we introduce Cracks in the Foundation (CiF), the largest and most detailed civil infrastructure (instance) segmentation dataset to date, comprising $\approx$150,000 high-resolution images meticulously curated over five years in collaboration with civil engineering experts. With the help of this unprecedented data source, we expose a blind spot of current visual AI: despite the advent of promptable Foundation Models (FMs) and Vision Language Models (VLMs), and despite the impressive abilities of today's specialised segmentation models, it turns out that dense image understanding in the built environment is nowhere near solved. Our evaluations indicate that even the most recent zero-shot FMs face significant challenges when deployed on real-world infrastructure and even the performance of specialised models with domain-specific supervision plateaus at $\approx$25% mAP. CiF establishes inspection of civil infrastructure, an elementary and seemingly easy perceptual task, as an open challenge that reveals fundamental weaknesses of present-day models trained predominantly on internet images, literally and figuratively highlighting cracks in the current foundation model paradigm.

Julius Erbach, Dominik Narnhofer, Andreas Dombos et al.

Flow-based latent generative models such as Stable Diffusion 3 are able to generate images with remarkable quality, even enabling photorealistic text-to-image generation. Their impressive performance suggests that these models should also constitute powerful priors for inverse imaging problems, but that approach has not yet led to comparable fidelity. There are several key obstacles: (i) the data likelihood term is usually intractable; (ii) learned generative models cannot be directly conditioned on the distorted observations, leading to conflicting objectives between data likelihood and prior; and (iii) the reconstructions can deviate from the observed data. We present FLAIR, a novel, training-free variational framework that leverages flow-based generative models as prior for inverse problems. To that end, we introduce a variational objective for flow matching that is agnostic to the type of degradation, and combine it with deterministic trajectory adjustments to guide the prior towards regions which are more likely under the posterior. To enforce exact consistency with the observed data, we decouple the optimization of the data fidelity and regularization terms. Moreover, we introduce a time-dependent calibration scheme in which the strength of the regularization is modulated according to off-line accuracy estimates. Results on standard imaging benchmarks demonstrate that FLAIR consistently outperforms existing diffusion- and flow-based methods in terms of reconstruction quality and sample diversity. Our code is available at https://inverseflair.github.io/.

Nando Metzger, Prune Truong, Goutam Bhat et al.

The growing demand for immersive 3D content calls for automated monocular-to-stereo video conversion. We present Elastic3D, a controllable, direct end-to-end method for upgrading a conventional video to a binocular one. Our approach, based on (conditional) latent diffusion, avoids artifacts due to explicit depth estimation and warping. The key to its high-quality stereo video output is a novel, guided VAE decoder that ensures sharp and epipolar-consistent stereo video output. Moreover, our method gives the user control over the strength of the stereo effect (more precisely, the disparity range) at inference time, via an intuitive, scalar tuning knob. Experiments on three different datasets of real-world stereo videos show that our method outperforms both traditional warping-based and recent warping-free baselines and sets a new standard for reliable, controllable stereo video conversion. Please check the project page for the video samples https://elastic3d.github.io.

Francesc Marti-Escofet, Benedikt Blumenstiel, Linus Scheibenreif et al.

Earth observation (EO) is crucial for monitoring environmental changes, responding to disasters, and managing natural resources. In this context, foundation models facilitate remote sensing image analysis to retrieve relevant geoinformation accurately and efficiently. However, as these models grow in size, fine-tuning becomes increasingly challenging due to the associated computational resources and costs, limiting their accessibility and scalability. Furthermore, full fine-tuning can lead to forgetting pre-trained features and even degrade model generalization. To address this, Parameter-Efficient Fine-Tuning (PEFT) techniques offer a promising solution. In this paper, we conduct extensive experiments with various foundation model architectures and PEFT techniques to evaluate their effectiveness on five different EO datasets. Our results provide a comprehensive comparison, offering insights into when and how PEFT methods support the adaptation of pre-trained geospatial models. We demonstrate that PEFT techniques match or even exceed full fine-tuning performance and enhance model generalisation to unseen geographic regions, while reducing training time and memory requirements. Additional experiments investigate the effect of architecture choices such as the decoder type or the use of metadata, suggesting UNet decoders and fine-tuning without metadata as the recommended configuration. We have integrated all evaluated foundation models and techniques into the open-source package TerraTorch to support quick, scalable, and cost-effective model adaptation.

Amaru Caceres Arroyo, Lea Bogensperger, Ahmed Allam et al.

Protein fitness optimization is challenged by a vast combinatorial landscape where high-fitness variants are extremely sparse. Many current methods either underperform or require computationally expensive gradient-based sampling. We present CHASE, a framework that repurposes the evolutionary knowledge of pretrained protein language models by compressing their embeddings into a compact latent space. By training a conditional flow-matching model with classifier-free guidance, we enable the direct generation of high-fitness variants without predictor-based guidance during the ODE sampling steps. CHASE achieves state-of-the-art performance on AAV and GFP protein design benchmarks. Finally, we show that bootstrapping with synthetic data can further enhance performance in data-constrained settings.

Mehmet Ozgur Turkoglu, Dominik J. Mühlematter, Alexander Becker et al.

Foundation models have become a dominant paradigm in machine learning, achieving remarkable performance across diverse tasks through large-scale pretraining. However, these models often yield overconfident, uncalibrated predictions. The standard approach to quantifying epistemic uncertainty, training an ensemble of independent models, incurs prohibitive computational costs that scale linearly with ensemble size, making it impractical for large foundation models. We propose Singular Value Ensemble (SVE), a parameter-efficient implicit ensemble method that builds on a simple, but powerful core assumption: namely, that the singular vectors of the weight matrices constitute meaningful subspaces of the model's knowledge. Pretrained foundation models encode rich, transferable information in their weight matrices. If the singular vectors are indeed meaningful (orthogonal) "knowledge directions". To obtain a model ensemble, we modulate only how strongly each direction contributes to the output. Rather than learning entirely new parameters, we freeze the singular vectors and only train per-member singular values that rescale the contribution of each direction in that shared knowledge basis. Ensemble diversity emerges naturally as stochastic initialization and random sampling of mini-batches during joint training cause different members to converge to different combinations of the same underlying knowledge. SVE achieves uncertainty quantification comparable to explicit deep ensembles while increasing the parameter count of the base model by less than 1%, making principled uncertainty estimation accessible in resource-constrained settings. We validate SVE on NLP and vision tasks with various different backbones and show that it improves calibration while maintaining predictive accuracy.