Bilevel Optimization Improves Auxiliary Learning for Medical Segmentation
A new method frames auxiliary learning as a bilevel optimization problem to make medical image segmentation more label-efficient.
Research Digest
Preprints and papers, rewritten in plain English.
WaveSpec adapts spectral coherence for iris attack detection
A new arXiv preprint proposes a wavelength-conditioned approach to improve multispectral iris presentation attack detection.
HiQGS: Spatiotemporal Querying for Feed-Forward Gaussian Reconstruction
A new approach jointly reasons over viewpoint and time to reconstruct Gaussian scenes from sparse monocular frames.
NeoWorld-3: Code-Defined Physical Worlds for Interactive 3D
A new framework treats 3D world reconstruction as a coding problem to enable physical interactivity, not just visual realism.
PACE-MoE: Adaptive Domain Sampling for Medical Visual Grounding
A new mixture-of-experts approach adaptively samples medical imaging domains and applies domain-specific LoRA to improve visual grounding across heterogeneous data.
Optimal Transport Framework for Frame Selection in Long Videos
A new arXiv paper proposes a principled optimal transport approach to choose frames for long-video understanding with multimodal LLMs.
Reusing AI Assurance Evidence: Validity Conditions and a Fail-Closed Gate
A new framework defines when prior governance checks can be safely reused in AI pipelines, tested on 714 public pipelines.
LiDAR Perception Must Balance Robustness and Efficiency
A new arXiv paper highlights the dual challenge of making LiDAR-based 3D detection robust to sensor corruption while staying within latency and memory limits.
PanoGraphGS Unifies Panoptic Segmentation and 3D Scene Graphs
A new arXiv paper proposes a unified representation that combines panoptic segmentation with hierarchical 3D scene graphs to support language-guided navigation.
VEOcc++ Maps Semantics Online and Refines Offline
A new arXiv paper proposes persistent voxel feature maps to unify online and offline semantic occupancy mapping for embodied agents.
RLSplat aims to make feed-forward 3D Gaussian splatting more resource-efficient
A new arXiv paper proposes resource-linear budget control to improve the practicality of single-pass 3D scene reconstruction.
Channel-Selective Fusion Network Aligns SAR and Optical Imagery for Object Detection
A new arXiv paper proposes a channel-selective interaction fusion network to better combine complementary SAR and optical imagery for remote sensing object detection.
World2HOSI: Multi-Expert Framework for Long-Horizon Human-Object-Scene Interaction
The framework combines specialized HOI and HSI models to address the challenge of generating long-horizon interactions that involve both objects and scenes.
Push-Pull Attentional Anchoring Targets Concept Erasure in Diffusion Models
A new arXiv paper proposes a push-pull attentional anchoring technique to remove unwanted concepts from pretrained diffusion models, addressing privacy and safety concerns.
RoboJudge uses MLLMs to check embodied video rollouts
A new framework proposes using multimodal large language models as judges to ensure generated videos are physically valid and faithful to task instructions.
Smart Glasses Pose New Challenge for Face Anti-Spoofing
A new study highlights how smart glass capture complicates face presentation attack detection, proposing decoupled representation learning.
Radar-Based Method Estimates 3D Human Pose for Smart Homes
A new approach uses millimeter-wave radar and a two-stage refinement pipeline to estimate 3D human poses indoors without relying on cameras.
PixelSR: Orthogonal Flow Decoupling for Real-World Super-Resolution
A new preprint proposes a pixel-space orthogonal flow decoupling method to address the long-standing seesaw effect in latent diffusion image super-resolution.