Computer Vision and Pattern Recognition
Covers image processing, computer vision, pattern recognition, and scene understanding. Roughly includes material in ACM Subject Classes I.2.10, I.4, and I.5.
sort pith recommended most recent
PatternGSL encodes panel boundaries, seams and stitch topology without templates; a VLM predicts it and deterministic rules decode to valid
· “PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments”
A black-box trigger reroutes infrared image-text alignment and carries over to captions and VQA.
· “QR-Structured Thermal Triggers for Targeted Semantic Attacks on Infrared Vision-Language Models”
Specialist reaches 0.92 precision on 100-image benchmark after training on 13k literature-mined fracture surfaces
· “Fine-tuning a vision-language model for fracture-surface morphology recognition”
Matching retrieval scope, visual inspection, and refusal to each question beats adding more context.
· “TAP-RAG: Task-Aware Policy Control for Long-Document Multimodal Question Answering”
RetiSEM organises variables into biological blocks and forbids edges, lowering structural error versus unconstrained models on synthetic tes
· “RetiSEM: Generalising Causal Models for Fragmented Biomedical Data”
Downsampled latent prediction and state-fusion decoding deliver 72 ms latency and 4.1 GiB memory on LIBERO and RoboTwin.
· “Light-WAM: Efficient World Action Models with State-Fusion Action Decoding”
MotionPDE solves a fluid-inspired equation under contrastive guidance to capture motion in unordered sequences and plugs into backbones with
· “Paving the Way for Point Cloud Video Representation Learning Using A PDE Model”
An SSM-Attention-SSM pipeline preserves local geometry while capturing global context for improved accuracy.
Multi-modal controls and step-by-step refinement raise identity retrieval rates over standard text-to-image diffusion in crime scenarios.
MSR aligns each vertebra rigidly then applies gated Mamba-Swin deformable correction before fusing into one hybrid field and releases an new
Monotone merge of dual signals plus mean alignment yields more reliable reliability estimates while preserving selective prediction trade-o
· “Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models”
TowerDataset supplies 22 fine classes and a fusion framework to handle long corridors and rare components in point cloud inspection.
Scale-aware aggregation and progressive widening of boundary attention maps improve accuracy on four benchmark datasets while keeping the U-
40k assets and EmissionGen baseline enable meshes to reproduce emission effects such as LEDs from reference images.
A 940k-frame simulated dataset transfers zero-shot to real sensors and cuts annotation needs to 5%.
· “SynFlow: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data”
Over 100K images across seven domains and three tasks test knowledge transfer for manufacturing quality checks.
· “CAD 100K: A Comprehensive Multi-Task Dataset for Car Related Visual Anomaly Detection”
CAMotion supplies annotated sequences with occlusion, blur and shape complexity to support broader algorithm testing.
· “CAMotion: A High-Quality Benchmark for Camouflaged Moving Object Detection in the Wild”
Street-legal attachment reaches 18 percent impersonation rate for under 100 dollars without altering plates
· “Street-Legal Physical-World Adversarial Rim for License Plates”
Local and global frequency penalties yield sharper, more consistent reconstructions from limited input images
Using pseudo ground-truth trajectories stabilizes training on fixed simulator data and raises route completion over imitation baselines.
A retrained segmentation model reads pre-destruction CORONA photos, confirming sites invisible on modern maps.
Pretrained on over a million hours of web video plus under 62 hours of unlabeled robot data, it enables pick-and-place planning without any
· “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning”
Cuts a recursive transformer's activation memory and speeds batch inference with just a 1.47-point accuracy drop.
· “MergeOver: Post-Training Token Merging for Recursive Vision Transformers”
Separating geometry from texture and painting with a diffusion renderer produces animatable 3D humans in minutes.
· “TGRHuman: Text-Guided Realistic 3D Human Generation via Diffusion Renderer”
A light head on a pretrained HPE model hits mAP 0.931 and trims triangulation error by up to 10.1%.
· “Hand Visibility Detector: Per-Keypoint Visibility Estimation for Hands”
REST turns reward-scored RL rollouts into distillation data, merging alignment and compression in one training stage.
· “RL-Native Distillation: Exploiting Scored Trajectories for Few-Step Image Generation”
Tiny pixel perturbation targets the attention mechanism of unified editors like FLUX.2, beating prior cloaks with less distortion.
· “VETO: Towards Protecting Images From Frontier AI Editing”
Sol-Attn routes, approximates, and computes in one pass — no retraining needed.
· “Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification”
A gated corrective-intent transfer beats robot-only and direct co-training on real failure starts.
· “EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration”
A single flow-matching model handles every observed–missing pair and preserves tumor detail.
· “Posterior Samplings are Missing Modalities Generators for Medical Image Translation”
Curating reliable image-text predictions and matching prototypes yields up to +12.2% DSC over fine-tuning on optic disc and lung benchmarks.
· “Memory-Supported Synergistic Adaptation for Training-Free Test-Time Medical Image Segmentation”
Swapping real footage into a teacher's memory cache during training gives streaming video models dense corrective targets with zero added推理
· “OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators”
A distilled video diffusion model sustains 720p/60fps interactive world generation across 20 scenarios, paired with an agentic harness for因果
Latent actions, dual transformers, and dream forcing close the gaps that break world-action models
· “ABot-M0.5: Unified Mobility-and-Manipulation World Action Model”
Direct 3D inference from point sets prunes redundancy and improves rendering efficiency over per-view methods.
· “PointSplat: Compact Gaussian Splatting via Human-Centric Prediction”
X-ray microtomography and refined algorithms recover full text from unopened ancient papyrus without damage.
· “Complete virtual unwrapping and reading of a rolled Herculaneum papyrus”
Initial neural-network estimate refined by optimization yields Strehl ratio 0.96 in 2-4 steps on real laser distortions.
· “Hybrid deep learning-based phase diversity method for wavefront reconstruction”
Camera pose injection, epipolar attention and distillation raise planning and risk scores on nuScenes and related benchmarks without auxilia
· “OmniSpace: Efficient Geometry Awareness for Autonomous Vehicles MLLMs”
A planner directs 2D-to-3D evidence collection and memory integration, raising benchmark performance without model retraining.
· “S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence”
Training-free STORM keeps 2D grid structure intact so selective scanning works after token reduction.
· “Spatial-Aware Reduction Framework: Towards Efficient and Faithful Visual State Space Models”
UV rendering plus diffusion textures produce natural patterns that stay effective as targets move and illumination varies.
· “VFACamou: View-Fused Adversarial Camouflage for Environment-Adaptive Physical Evasion”
Model reaches 85 percent F1 on expanded forest benchmark and holds performance on unseen sites
ROGLE mines pseudo region-sentence pairs from existing data to add local alignment, raising accuracy on detailed natural-language queries.
Geometry-guided flow places motion into the noise space, preserving diffusion priors for consistent trajectories.
AssetGen controls polygon count while baking normals and textures so the output runs in real time without further editing.
· “AssetGen: Deployable 3D Asset Generation at Interactive Speed”
Timestep-aware non-incremental aggregation fixes magnitude growth and gradient decay on ImageNet generation
· “Rethinking Cross-Layer Information Routing in Diffusion Transformers”
Progressive splatting with depth warping and semantic synthesis keeps views consistent at large magnifications.
UniPPTBench experiments show performance varies sharply by input setting and that visual appeal scores do not ensure content grounding or跨源s
· “UniPPTBench: A Unified Benchmark for Presentation Generation Across Diverse Input Settings”
X-Shift redirects heatmaps to irrelevant areas on CLIP models while outputs stay stable across datasets.
Radiology text prompts plus spectral cues enable joint type and severity prediction on single and mixed degradations, beating baselines.
· “CT-DegradBench: A Physics-Informed Benchmark for CT Degradation Detection and Severity Estimation”
A single switch from structure-biased to texture-biased control inside one diffusion process improves perceptual quality while keeping align
· “LPH-VTON: Resolving the Structure-Texture Dilemma of Virtual Try-On via Latent Process Handover”
Advantage routing, gradient surgery, and region reweighting fix inconsistencies that plague vanilla RL in joint diffusion models.
· “OmniNFT: Modality-wise Omni Diffusion Reinforcement for Joint Audio-Video Generation”
Coarse-to-fine modules using one reference view achieve zero-shot correction of blur and distortions across backbones and tasks.
· “UniFixer: A Universal Reference-Guided Fixer for Diffusion-Based View Synthesis”
By extracting latent actions from semantic, depth, and flow transitions, MoLA improves success and consistency over direct use of generated帧
· “From Imagined Futures to Executable Actions: Mixture of Latent Actions for Robot Manipulation”
Hierarchical visual questions turn off-the-shelf models into safety classifiers across 20 cities, outperforming classification methods.
· “Urban Risk-Aware Navigation via VQA-Based Event Maps for People with Low Vision”
LDDR selects non-redundant frames query-aware and gives them higher resolution, delivering 2.5-point gains under token limits while running
· “LDDR: Linear-DPP-Based Dynamic-Resolution Frame Sampling for Video MLLMs”
250 prompts with expected outcomes and 13 laws from mechanics to optics support law-by-law checks
· “PhyGround: Benchmarking Physical Reasoning in Generative World Models”
Two modules reshape cross-attention to correct actor mismatches and target errors without any model retraining.
· “SocialDirector: Training-Free Social Interaction Control for Multi-Person Video Generation”
Probabilistic scores capture annotator differences while regularization stabilizes knapsack selection for competitive results in one pass.
· “Uncertainty-Aware and Decoder-Aligned Learning for Video Summarization”
Geometrically constrained optimal transport creates synthetic angiograms that raise downstream detector accuracy on benchmarks.
· “Geometrically Constrained Stenosis Editing in Coronary Angiography via Entropic Optimal Transport”
Global appearance and local dynamic structures are fused into a graph that lets the world model predict states for physically consistent 4D
· “ST-Gen4D: Embedding 4D Spatiotemporal Cognition into World Model for 4D Generation”
The local-support kernel replaces exponential smoothing in implicit moving least squares to retain fine geometry and enable direct end-toend
· “High-Fidelity Surface Splatting-Based 3D Reconstruction from Multi-View Images”
By organizing motion codes according to gesture semantics and applying reference prompts, the system produces motions faithful to both word,
Transcribing where-what-how-how-many details from ground-truth labels yields better accuracy than large language model methods on a new Gaza
Cross-modal mapping and dual prompts let models spot 3D defects in unseen categories without target training data.
Frozen backbones plus nearest-centroid classification deliver high accuracy and stable binary transfer across three public datasets.
· “Few-Shot Learning Pipeline for Monkeypox Skin Disease Classification Using CNN Feature Extractors”
Cascaded command and parameter diffusion with custom transition matrices outperforms continuous and autoregressive baselines on DeepCAD.
· “Computer-Aided Design Generation by Cascaded Discrete Diffusion Model”
Counterfactual granule supervision separates high-frequency details to enhance generalization and stability.
· “SpecPL: Disentangling Spectral Granularity for Prompt Learning”
TrajRAG converts past episodes into compact topo-polar trajectories and retrieves matching ones to improve zero-shot waypoint decisions.
· “TrajRAG: Retrieving Geometric-Semantic Experience for Zero-Shot Object Navigation”
A fixed set of 500 examples makes existing LLMs write reliable 3D modeling scripts on ordinary hardware.
· “BlenderRAG: High-Fidelity 3D Object Generation via Retrieval-Augmented Code Synthesis”
By structuring visuals as semantic graphs and aligning them topologically with past scenarios, the system handles rare situations more-reli
· “VLADriver-RAG: Retrieval-Augmented Vision-Language-Action Models for Autonomous Driving”
RSCNet uses cross-source guidance for spectral selection and adaptive fusion to outperform prior methods with lower complexity on benchmark
· “Representative Spectral Correlation Network for Multi-source Remote Sensing Image Classification”
Models trained on the resulting dataset improve captioning accuracy and step localization while generalizing across camera perspectives.
· “DenseStep2M: A Scalable, Training-Free Pipeline for Dense Instructional Video Annotation”
Treating ripening as a continuous distribution rather than discrete classes improves reliability when annotator labels contain boundary fuzz
· “FruitProM-V2: Robust Probabilistic Maturity Estimation and Detection of Fruits and Vegetables”
OmniShotCut uses a Transformer with shot queries to jointly model ranges and relations, trained on fully synthetic transitions for exact, sc
· “OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer”
Lightweight adapters factorize instructions from spatial regions and let a learned predictor supply the mask, beating both mask-free and or-
· “Edit Where You Mean: Region-Aware Adapter Injection for Mask-Free Local Image Editing”
By scoring its own candidate molecules against validity metrics, the model beats prior methods on real patent and literature images with far
· “COMO: Closed-Loop Optical Molecule Recognition with Minimum Risk Training”
A new calibration technique plus an objectness token in CLIP deliver reliable pseudo labels and set new performance records on COCO and LVIS
· “Exploring Hierarchical Consistency and Unbiased Objectness for Open-Vocabulary Object Detection”
Sparse Forcing adds a native trainable sparsity mechanism and PBSA kernel to autoregressive diffusion video models, yielding higher VBench…
Aerial images have objects of varying sizes and complex scenes, so the new framework with mixed attention and staged decoding succeeds on a
· “RefAerial: A Benchmark and Approach for Referring Detection in Aerial Images”
