OMR System for String Quartet PDF Scores
A new optical music recognition pipeline uses SegNet segmentation and convolutional processing to digitize string quartet scores from PDFs.
Research Digest
Preprints and papers, rewritten in plain English.
BodyBridge Targets Realistic VRU Translation in Driving Simulators
A new arXiv paper proposes grounded body layers to improve how simulation-to-real video models represent pedestrians and cyclists.
Class-Conditioned Guidance Keeps Medical AI Focused on Relevant Regions
A new arXiv paper proposes class-conditioned explanation guidance to stop medical imaging models from relying on unrelated image regions.
Spatially Localized Neural Operators for Efficient XR Predistortion
A new family of compositional, spatially localized neural operators could make XR display distortion correction more computationally efficient.
Spora: A Unified Dataset and Platform for Spatial Proteomics
A new arXiv preprint introduces spora, a unified multimodal dataset and platform designed to support spatial proteomics research at single-cell resolution.
CRISP: A Curriculum-Based Method for Low-Light Video Enhancement
A new approach combines a progressive illumination curriculum with semantic priors to help low-light video enhancement balance brightness, detail, and temporal coherence.
Hexagonal Basis Projectors Proposed for CT Reconstruction
A new arXiv preprint argues that representing CT images on a hexagonal lattice could make model-based iterative reconstruction more efficient.
Unified Multimodal Models Take on Fine-Grained Visual Understanding
A new arXiv paper highlights the challenge of teaching vision-language models to discern subtle local attributes and spatial relations.
Unified Latent Flow for Coherent Human-Object Interaction Generation
A new method treats human and object motion as a single latent flow to address the coherence problem in text-conditioned 3D interaction generation.
BlenderWorld Benchmark Tests AI Agents on 3D Creation from Real-World References
A new arXiv paper introduces BlenderWorld, a benchmark for evaluating multimodal LLM agents on 3D modeling tasks grounded in real-world references.
Satellite Imagery and Machine Learning Track Rooftop Solar Growth
A new arXiv paper proposes using satellite imagery and machine learning to keep pace with the rapid deployment of rooftop solar panels on industrial buildings.
Hierarchical Video-Action Model Targets Vision-Language Navigation
A new arXiv preprint proposes HAVANA, a hierarchical video-action model for handling both long-horizon and local planning in vision-language navigation.
Scale-Blind Registration Aligns Preoperative Liver Models to Intraoperative Surfaces
A new arXiv paper describes a landmark-free, scale-blind method for aligning preoperative liver models with intraoperative surfaces to support augmented-reality guidance in laparoscopic surgery.
Query-Aware Token Pruning Aims to Trim High-Resolution LVLM Overhead
A new arXiv paper tackles the computational burden of thousands of vision tokens in high-resolution large vision-language models.
Counterfactual Cardiac Anatomy via Causal Neural Distance Fields
A new arXiv preprint proposes a causal spatio-temporal framework to predict how a patient's heart would change under altered clinical conditions.
OrTeN: Neural Inverse Graphics for Joint Terrain and Camera Trajectory
A new arXiv paper proposes OrTeN, a neural inverse graphics approach that jointly reconstructs digital terrain maps and line-scan camera trajectories for planetary exploration.
Bitstream-Corrupted Image Restoration Gets a Native Restorer
A new method, BiNar, targets the difficult task of reconstructing images from corrupted compressed bitstreams.
AI method recovers erased text in palimpsest manuscripts
A new approach combines generative text modeling with language-guided visual attention to uncover overwritten writing in reused manuscripts.