Tracking Objects Through Visibility Gaps in 4D Scene Graphs
A new method addresses the challenge of maintaining object identity when objects disappear inside opaque carriers and reappear later.
Research Digest
Preprints and papers, rewritten in plain English.
AnomalyJev proposes claim-level checks for anomaly explanations
A new arXiv preprint introduces AnomalyJev, a framework that evaluates multimodal LLM anomaly-detection explanations by verifying individual claims.
Less Is More: New Approach to Composed Image Retrieval
A paper on zero-shot composed image retrieval suggests that concise retrieval and critic-based re-ranking may improve on methods that rely on large vision-language models.
SGLDet: Spectral-Aware Modeling for Aerial Object Detection
A new arXiv preprint proposes a spectral-aware global-local dynamic model to improve multispectral object detection in complex aerial scenes.
Video Unlearning Leaks via Frame Redundancy, Not Trajectory
A new arXiv study shows that privacy leaks in video class unlearning come from redundant frame information rather than motion trajectories, exposing blind spots in current evaluation methods.
NeRF Geometry Gets Regularization via Cross-Ray Consistency
A new self-supervised technique aims to reduce the noisy geometry that plain NeRF models produce, without changing the core architecture.
Matrix-Game 4.0 Roadmap: From Controllable Videos to Playable Worlds
A new arXiv paper outlines the requirements for turning video models into real-time playable game worlds.
Medical VLMs Evaluated on Clean Inputs Despite Clinical Evidence Loss
A new preprint warns that medical vision-language models are typically judged on pristine images, ignoring the progressive evidence loss common in real clinical settings.
Real In, Real Out: Scene Text Editing with Real Data
A new paper asks whether scene text editing can be done using only real images, sidestepping the usual need for paired before-and-after training data.
Point-based 3D VLM targets semantic leakage from bounding boxes
A new arXiv preprint argues that bounding-box prompts in 3D vision-language benchmarks cause semantic leakage and proposes a point-based alternative.
Configurable Perception in World Models Targets Noisy Observations
A new arXiv preprint argues that world models need configurable perception to filter task-irrelevant noise from real-world observations.
AI extracts food-package evidence for South Africa's R3337 draft
A new approach uses region-guided extraction and abstention-aware screening to apply draft R3337 warning criteria to food-package photos.
Causal State Evolution Proposed for 4D Scene Reconstruction
A new arXiv paper argues that dynamic 4D reconstruction should treat visual geometry as a causally evolving state rather than using unordered temporal aggregation.
SplatFlux: Feed-Forward 4D Reconstruction and Generation from Video Latents
A new arXiv preprint proposes a unified feed-forward framework that maps video latents to 4D Gaussian splats for both reconstruction and conditional generation.
Psychophysics-inspired model reads emotions in oddly satisfying videos
A new multimodal LLM uses psychophysical principles to recognise emotional responses in videos that lack explicit emotional cues.
Generative Deception Could Force AI Attackers to Waste Energy
A new framework proposes using generative deception to make language-model attackers spend more computational resources, potentially shifting the defender-attacker cost balance.
Looped Transformer Predicts Gene Expression from Histology Images
A new method called HistoLoop uses iterative contextual inference to predict spatial gene expression from H&E-stained tissue images.
Can one encoder unify diverse geospatial foundation models?
A new arXiv study asks whether a single encoder can combine the strengths of many geospatial foundation models into one consistent representation.