A5084415727- Generative Adversarial Networks and Image Synthesis
- Advanced Image and Video Retrieval Techniques
- Human Pose and Action Recognition
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Image Retrieval and Classification Techniques
- Advanced Vision and Imaging
- Advanced Neural Network Applications
- Computer Graphics and Visualization Techniques
- Advanced Image Processing Techniques
- Video Surveillance and Tracking Methods
- Face recognition and analysis
- Image Processing and 3D Reconstruction
- 3D Surveying and Cultural Heritage
- Anomaly Detection Techniques and Applications
- Robotics and Sensor-Based Localization
- Explainable Artificial Intelligence (XAI)
- Human Motion and Animation
- Video Analysis and Summarization
- Medical Imaging Techniques and Applications
- Aesthetic Perception and Analysis
- Image Enhancement Techniques
- Hand Gesture Recognition Systems
- Image and Signal Denoising Methods
- Remote-Sensing Image Classification
Ludwig-Maximilians-Universität München
2021-2025
LMU Klinikum
2023-2025
Heidelberg University
2012-2022
Heidelberg University
2012-2021
University of California, Berkeley
2009
ETH Zurich
2005-2007
University of Bonn
2003
By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on data and beyond. Additionally, their formulation allows for guiding mechanism to control generation without retraining. However, since these typically operate directly in pixel space, optimization powerful DMs often consumes hundreds GPU days inference is expensive due evaluations. To enable DM training limited computational...
Designed to learn long-range interactions on sequential data, transformers continue show state-of-the-art results a wide variety of tasks. In contrast CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness CNNs with expressivity enables model and thereby synthesize (i) use context-rich vocabulary image...
The brain exhibits limited capacity for spontaneous restoration of lost motor functions after stroke. Rehabilitation is the prevailing clinical approach to augment functional recovery, but scientific basis poorly understood. Here, we show nearly full recovery skilled forelimb in rats with large strokes when a growth-promoting immunotherapy against neurite growth-inhibitory protein was applied boost sprouting new fibers, before stabilizing newly formed circuits by intensive training. In...
Artists rarely paint in a single style throughout their career. More often they change styles or develop variations of it. In addition, artworks different and even within one depict real content differently: while Picasso's Blue Period displays vase blueish tone but as whole, his Cubist works deconstruct the object. To produce artistically convincing stylizations, transfer models must be able to reflect these changes variations. Recently many have aimed improve task, neglected address...
Large intra-class variation is the result of changes in multiple object characteristics. Images, however, only show superposition different variable factors such as appearance or shape. Therefore, learning to disentangle and represent these characteristics poses a great challenge, especially unsupervised case. Moreover, large articulation calls for flexible part-based model. We present an approach disentangling shape by parts consistently over all instances category. Our model representation...
Capturing photographs with wrong exposures remains a major source of errors in camera-based imaging. Exposure problems are categorized as either: (i) overexposed, where the camera exposure was too long, resulting bright and washed-out image regions, or (ii) underexposed, short, dark regions. Both under- overexposure greatly reduce contrast visual appeal an image. Prior work mainly focuses on underexposed images general enhancement. In contrast, our proposed method targets both over-...
By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on data and beyond. Additionally, their formulation allows for guiding mechanism to control generation without retraining. However, since these typically operate directly in pixel space, optimization powerful DMs often consumes hundreds GPU days inference is expensive due evaluations. To enable DM training limited computational...
Abstract Nerve injury leads to chronic pain and exaggerated sensitivity gentle touch (allodynia) as well a loss of sensation in the areas which injured non-injured nerves come together 1–3 . The mechanisms that disambiguate these mixed paradoxical symptoms are unknown. Here we longitudinally non-invasively imaged genetically labelled populations fibres sense noxious stimuli (nociceptors) (low-threshold afferents) peripherally skin for longer than 10 months after nerve injury, while...
Abstract The field of visual computing is rapidly advancing due to the emergence generative artificial intelligence (AI), which unlocks unprecedented capabilities for generation, editing, and reconstruction images, videos, 3D scenes. In these domains, diffusion models are AI architecture choice. Within last year alone, literature on diffusion‐based tools applications has seen exponential growth relevant papers published across computer graphics, vision, communities with new works appearing...
Detecting abnormalities in video is a challenging problem since the class of all irregular objects and behaviors infinite thus no (or by far not enough) abnormal training samples are available. Consequently, standard setting to find without actually knowing what they because we have been shown examples during training. However, although data does define an abnormality looks like, main paradigm this field directly search for individual local patches or image regions independent another. To...
Learning the embedding space, where semantically similar objects are located close together and dissimilar far apart, is a cornerstone of many computer vision applications. Existing approaches usually learn single metric in space for all available data points, which may have very complex non-uniform distribution with different notions similarity between objects, e.g. appearance, shape, color or semantic meaning. Approaches learning distance often struggle to encode types relationships do not...
Estimating dense visual correspondences between objects with intra-class variation, deformations and background clutter remains a challenging problem. Thanks to the breakthrough of CNNs there are new powerful features available. Despite their easy accessibility great success, existing semantic flow methods could not significantly benefit from these without extensive additional training. We introduce novel method for matching pre-trained CNN which is based on convolutional feature pyramids...
Metric learning seeks to embed images of objects such that class-defined relations are captured by the embedding space. However, variability in is not just due different depicted object classes, but also depends on other latent characteristics as viewpoint or illumination. In addition these structured properties, random noise further obstructs visual interest. The common approach metric enforce a representation invariant under all factors ones contrast, we propose explicitly learn shared and...
Style transfer has recently received a lot of attention, since it allows to study fundamental challenges in image understanding and synthesis. Recent work significantly improved the representation color texture computational speed resolution. The explicit transformation content has, however, been mostly neglected: while artistic style affects formal characteristics an image, such as color, shape or texture, also deforms, adds removes details. This paper explicitly focuses on content-and...
Current neuromodulatory strategies to enhance motor recovery after stroke often target large brain areas non-specifically and without sufficient understanding of their interaction with internal repair mechanisms. Here we developed a novel therapeutic approach by specifically activating corticospinal circuitry using optogenetics strokes in rats. Similar neuronal growth-promoting immunotherapy, optogenetic stimulation together intense, scheduled rehabilitation leads the restoration lost...
Neural networks have greatly boosted performance in computer vision by learning powerful representations of input data. The drawback end-to-end training for maximal overall are black-box models whose hidden lacking interpretability: Since distributed coding is optimal latent layers to improve their robustness, attributing meaning parts a feature vector or individual neurons hindered. We formulate interpretation as translation onto semantic concepts that comprehensible the user. mapping...
Exemplar learning is a powerful paradigm for discovering visual similarities in an unsupervised manner. In this context, however, the recent breakthrough deep could not yet unfold its full potential. With only single positive sample, great imbalance between one and many negatives, unreliable relationships most samples, training of Convolutional Neural networks impaired. Given weak estimates local distance we propose optimization problem to extract batches samples with mutually consistent...
There have been many successful implementations of neural style transfer in recent years. In most these works, the stylization process is confined to pixel domain. However, we argue that this representation unnatural because paintings usually consist brushstrokes rather than pixels. We propose a method stylize images by optimizing parameterized instead pixels and further introduce simple differentiable rendering mechanism.Our approach significantly improves visual quality enables additional...
Designed to learn long-range interactions on sequential data, transformers continue show state-of-the-art results a wide variety of tasks. In contrast CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness CNNs with expressivity enables model and thereby synthesize (i) use context-rich vocabulary image...
Autoregressive models and their sequential factorization of the data likelihood have recently demonstrated great potential for image representation synthesis. Nevertheless, they incorporate context in a linear 1D order by attending only to previously synthesized patches above or left. Not is this unidirectional, bias attention unnatural images as it disregards large parts scene until synthesis almost complete. It also processes entire on single scale, thus ignoring more global contextual...