Institution profile

University of Würzburg

Academic institutioneurope · de
Official website
Research library253linked papers
Opportunities0open roles
Selected work

Representative Papers

Rip Current Segmentation: A Novel Benchmark and YOLOv8 Baseline Results

Jun 01, 20232023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

This paper addresses the challenging problem of automated rip current detection in beach environments. We formally introduce the task of instance-level rip current segmentation—a novel formulation for fine-grained, pixel-accurate identification. To support this task, we construct the first high-quality multimodal benchmark dataset, comprising 2,466 images with polygon annotations and 17 drone-captured video sequences (24K frames) with frame-level mask annotations. Methodologically, we adapt the YOLOv8 instance segmentation architecture into a lightweight, real-time framework capable of processing both static images and video streams. Our optimized YOLOv8-nano variant achieves 88.94% mAP₅₀ on the validation set and 81.21% macro-average precision on test videos. All code, pre-trained models, and annotated data are publicly released. This work establishes a deployable baseline—suitable for edge devices—for intelligent, real-time rip current monitoring, thereby advancing both benchmarking and practical deployment in coastal safety applications.

18 citationsRead paper

RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report

Jun 02, 2025

This work addresses the inverse ISP problem: reconstructing high-fidelity RAW sensor data from metadata-free smartphone sRGB images—bypassing the irreversibility inherent in conventional ISP modeling. We introduce the first large-scale, end-to-end sRGB→RAW inverse ISP benchmark and establish a camera-agnostic reconstruction paradigm. Our method employs a conditional generative deep neural network, incorporating physics-inspired constraints and multi-scale perceptual loss to significantly enhance cross-device generalization. Evaluated on a benchmark with over 150 participating teams, our approach sets new state-of-the-art scores across PSNR, SSIM, and LPIPS. Reconstructed RAW images are rigorously validated via forward ISP rendering and real-sensor response simulation, demonstrating strong fidelity and consistency. This work opens new avenues for low-level vision modeling and synthetic RAW data generation.

12 citationsRead paper

NTIRE 2025 Challenge on Video Quality Enhancement for Video Conferencing: Datasets, Methods and Results

May 25, 2025

Real-world video conferencing suffers from low illumination, color distortion, prominent noise, and blurred details. Method: This paper proposes an end-to-end spatiotemporal lightweight Video Quality Enhancement (VQE) model that jointly optimizes illumination equalization, color fidelity, noise suppression, and sharpness enhancement. It introduces the first multi-objective VQE benchmark tailored to authentic meeting scenarios and incorporates a differentiable Video Quality Assessment (VQA) model to guide end-to-end training. A dual-track evaluation framework—combining objective metrics with crowdsourced subjective assessment—is also designed. Contribution/Results: In a public competition with 91 participating teams and 10 valid submissions, the top-performing method achieves significant improvements in key perceptual metrics—including LPIPS, NIQE, and MOS—demonstrating the feasibility and practicality of achieving studio-grade visual quality under low-bandwidth constraints.

9 citationsRead paper

Efficient Real-World Deblurring using Single Images: AIM 2025 Challenge Report

Oct 14, 2025

This work addresses efficient single-image deblurring for real-world scenarios under strict lightweight constraints: <5 M parameters and <200 GMACs. Leveraging our newly introduced RSBlur dataset—collected via a dual-camera setup and containing paired sharp-blurry images—we propose a lightweight convolutional backbone, a channel-spatial collaborative attention module, and a multi-stage feature recalibration mechanism to achieve high-fidelity restoration with minimal computational overhead. On the RSBlur test set, our method achieves 31.1298 dB PSNR, setting the new state-of-the-art among all approaches satisfying the specified efficiency constraints. Its feasibility and scalability are further validated by four independent participating teams in a benchmarking challenge. To the best of our knowledge, this is the first study to systematically define, construct, and empirically validate a practical lightweight benchmark for real-world image deblurring—bridging the gap between algorithmic performance and edge-device deployment.

3 citationsRead paper

NTIRE 2025 Challenge on Short-form UGC Video Quality Assessment and Enhancement: KwaiSR Dataset and Study

Apr 21, 2025

This paper addresses the low image quality of user-generated content (UGC) on short-video platforms and the poor generalization of existing single-image super-resolution (SR) methods. To bridge this gap, the authors introduce KwaiSR—the first benchmark dataset tailored to real-world UGC scenarios—comprising 1,800 synthetically generated LR-HR image pairs and 1,900 authentic low-quality images selected via KVQ’s quality assessment model, enabling dual-domain co-modeling of synthetic ground truth and realistic degradation distributions. Leveraging KwaiSR, the authors organized the NTIRE 2025 Challenge on Short-Format UGC Video Quality Assessment and Enhancement, attracting over 30 participating teams. Extensive experiments reveal substantial performance degradation of mainstream SR methods on UGC data, highlighting critical challenges in realistic degradation modeling, quality-aware image selection, and cross-domain generalization. KwaiSR thus establishes a foundational data resource and a new research paradigm for UGC image enhancement.

3 citationsRead paper
Recent publications

Latest Papers