LatticeRank: Native-Grid Segmentation That Respects Voxel Spacing
A new ranking-based consistency method lets 3D medical image segmentation run on native voxel grids, avoiding the distortions of resampling.
Research Digest
Preprints and papers, rewritten in plain English.
OV-Foam Study Examines Feature Lifting in 3D Gaussian Splatting
New research shows how overlapping primitives in 3D Gaussian splatting complicate the standard backprojection method used for open-vocabulary segmentation.
LoRA-Adapted Brain MRI Models: Similar Performance, Different Mechanisms
A new arXiv study suggests that vision foundation models adapted for brain MRI can match in performance while diverging in how they work internally.
PointForward V2 Improves Feedforward 3D Reconstruction of Driving Scenes
A new arXiv preprint extends the PointForward model with point-aligned interaction to better fuse multi-view, multi-frame observations into a coherent 3D representation.
Offline Calibration Improves Cache Compression for Video Generation
A new approach uses offline calibration to make token merging more effective for reducing memory in autoregressive video generation.
Attractive Faces Can Skew Text-to-Image Reward Models and Leaderboards
New research highlights a previously underexplored bias in human-preference-based evaluation of text-to-image models.
Multi-agent framework refines text-to-motion generation conversationally
A new arXiv paper proposes a self-evolving multi-agent system that lets users refine generated motion through conversation instead of regenerating from scratch.
GAMF-Net Fuses Smartphone Images and Clinical Data for Skin Lesion Classification
A new multimodal model aims to improve smartphone-based skin lesion classification by combining images with medical history and symptoms.
Unified RAW Exposure Correction and Restoration
A new arXiv paper proposes processing RAW sensor data with linear illumination and frequency-adaptive enhancement to recover detail lost to incorrect exposure.
WildAIGC-Bench: Benchmarking AI Image Detectors in the Wild
A new benchmark evaluates AI image detectors on images actually circulating online, rather than only on clean test sets.
Benchmark Targets Code-Based Video Generation Across Three Task Types
A new benchmark, Code2Video-Compass, aims to standardize evaluation of code-guided video generation for text-, image-, and video-to-video tasks.
CountVideoBench: New Benchmark for Counting Biases in Video-Language Models
A new benchmark aims to expose how video-language models are biased when counting objects in video.
Multi-agent framework automates ML for medical imaging
A new arXiv preprint proposes an evidence-based multi-agent system to handle the difficult work of model design in medical image analysis.
ABACUS Benchmark Exposes Counting Failures in Image Editing Models
A new benchmark for instruction-based image editing shows that models often miscount objects even when they successfully preserve the original scene.
RACE: Training Multimodal Search Without Visual Shortcuts
A new preprint argues that synthetic search tasks let agents ignore visual evidence, and proposes RACE to force genuine multimodal grounding.
Text Prompts May Not Actually Guide Vision-Language Trackers
New research introduces a degradation-indexed framework to test whether textual descriptions truly alter tracking decisions when visual evidence worsens.
Gated Prototype Transfer Targets Zero-Support Medical VLMs
A new method aims to keep medical vision-language models usable when rare classes have no labeled examples during adaptation.
Privacy-Preserving AI for Real-Time ECG Monitoring
An arXiv preprint argues that ECG monitoring's shift to distributed healthcare AI demands end-to-end privacy protection.