outlier rejection

Image matching faces challenges in robustness and accuracy under complex scenes, particularly for visual localization and 3D reconstruction. To address this, we systematically reformulate the conventional multi-stage pipeline and propose the first dual-dimensional taxonomy—aligned with the detection-description-matching-geometric-estimation workflow—to uniformly evaluate twelve deep matching paradigms across pose estimation, homography estimation, and visual localization. Our method integrates differentiable geometric solvers, end-to-end trainable architectures, contrastive/self-supervised feature learning, and robust optimization modules into a standardized benchmark. Extensive experiments reveal fundamental trade-offs among sparse, semi-dense, and dense matching strategies—as well as pose regression paradigms—in terms of accuracy, robustness, and efficiency. The study identifies key open challenges and delineates principled directions for next-generation matching frameworks.

Industrial anomaly detection suffers from insufficient robustness against adversarial attacks—particularly because models are trained exclusively on normal data, rendering them incapable of identifying maliciously crafted anomalies, and because unsupervised settings lack reliable adversarial objectives. To address this, we propose a robust training framework leveraging pseudo-anomalous sample construction and contrastive learning. We introduce the novel concept of “opposite sample pairs” and integrate an adversarial separation strategy to suppress interference from pseudo-negative samples, jointly optimizing inter-class discriminability and intra-class compactness to sharpen decision boundaries. Our method unifies pseudo-anomaly generation, self-supervised representation learning, and robust loss design. It achieves state-of-the-art performance on multiple benchmarks under both clean and adversarial conditions, with an average robust detection accuracy improvement of 26.1%.

Existing anomaly detection (AD) and localization (AL) methods rely exclusively on normal samples for training, resulting in poor adversarial robustness and insufficient reliability for safety-critical medical and industrial applications. To address this, we propose PatchGuard—a ViT-based framework introducing the first foreground-aware pseudo-anomaly generation paradigm, supervised by localization masks, and theoretically establishing an intrinsic connection between ViT’s self-attention mechanism and adversarial robustness. Furthermore, we design a novel adversarial loss function that jointly optimizes localization accuracy and robustness. Evaluated on benchmark industrial and medical datasets, PatchGuard improves AD and AL performance under adversarial attacks by 53.2% and 68.5%, respectively, while maintaining state-of-the-art accuracy in clean (non-adversarial) settings.

Existing correspondence matching methods—spanning stereo matching, optical flow, and feature matching—are typically task-specific, relying on dedicated architectures or fine-tuning. This hinders generalization across tasks and domains. Method: We propose PanMatch, the first universal foundation model for multi-frame correspondence matching. It introduces a unified 2D displacement estimation framework, leverages large vision models (LVMs) as generic feature extractors, and incorporates a lightweight feature transformation pipeline to enable zero-shot cross-task and cross-domain transfer. PanMatch is pre-trained on a large-scale, cross-domain dataset comprising nearly 1.8 million samples. Contribution/Results: PanMatch achieves shared-weight inference across all three tasks. Experiments show it outperforms UniMatch and Flow-Anything in cross-task evaluation, matches state-of-the-art task-specific methods, and demonstrates significantly improved zero-shot robustness under challenging conditions—including rainy scenes and satellite imagery.

Pixel-level visual anomaly detection fails in architectural environments due to dynamic camera pose changes, undermining conventional supervised and unsupervised approaches. Method: We formulate Scene AD—a novel unsupervised anomaly detection task requiring only unlabeled normal images and test images—enabling cross-view, multi-object scene localization. We propose OmniAD, an improved reverse distillation framework integrating adaptive NeRF-based view synthesis and pose-estimation-guided data augmentation to jointly model unsupervised feature reconstruction and discrepancy. Contribution/Results: Evaluated on two newly constructed benchmarks—ToyCity (multi-object) and MAD (single-object)—OmniAD achieves a 40% improvement in pixel-level detection performance over state-of-the-art unsupervised methods, significantly enhancing practicality and robustness for open-scene anomaly detection.

Existing 3D anomaly detection methods struggle to effectively model the relationships between multi-view and multi-modal features, limiting their anomaly localization accuracy. To address this challenge, this work proposes ModMap, a novel framework that jointly models multi-view and multi-modal information by integrating cross-modal feature mapping with a view-aware feature modulation mechanism for the first time. The approach further introduces an omnidirectional ensemble training strategy to generate comprehensive multi-view anomaly scores. Additionally, the project releases a high-resolution depth encoder tailored for industrial applications. Evaluated on the SiM3D benchmark, ModMap significantly outperforms current state-of-the-art methods, achieving the best reported performance in both 3D anomaly detection and segmentation.

This work addresses the limitations of existing reconstruction-based unsupervised anomaly detection methods, which struggle to capture subtle anomalies and precisely delineate ambiguous boundaries, particularly in multi-class settings. To overcome these challenges, the study introduces a novel pseudo cross-modal matching framework that treats the input and reconstructed images as complementary modalities. By leveraging an attention-guided cross-modal matching mechanism and an adaptive frequency-aware fusion module, the method explicitly models anomaly-relevant patterns and enhances fine-grained detail perception through multi-scale high-frequency component coupling. Extensive experiments demonstrate state-of-the-art performance on the MVTec-AD, VisA, and MPDD benchmarks, significantly outperforming current best approaches and exhibiting superior capability in both complex anomaly detection and precise localization.

This work addresses the limited sensitivity of reconstruction-based paradigms to anomalous regions in unsupervised anomaly detection by formulating flow matching as a geometric dynamical process. The authors propose a “velocity mismatch” mechanism that detects anomalies by comparing the model-predicted velocity with the local geometric velocity of test images along affine trajectories, eliminating the need for test-time optimization or additional calibration. Theoretical analysis reveals that this mismatch decomposes into a denoising term and a Fisher divergence term, inspiring a multi-path aggregation strategy to enhance robustness. Extensive experiments demonstrate that the proposed method significantly outperforms existing reconstruction-based and flow-matching approaches on MVTec-AD and VisA benchmarks, achieving state-of-the-art performance in both pixel-level anomaly localization and image-level anomaly scoring.

This work addresses the challenge of feature inconsistency in multi-view, multi-modal industrial defect detection caused by viewpoint variations and modality discrepancies. To achieve physically consistent cross-view and cross-modal feature representations, the authors propose a unified framework named SGANet that jointly models semantic and geometric alignment. The method innovatively integrates a Selective Cross-view Feature Refinement Module (SCFRM), Semantic-Structural Patch Alignment (SSPA), and Multi-View Geometry Alignment (MVGA) to cooperatively optimize feature interaction and consistency. Extensive experiments on the SiM3D and Eyecandies datasets demonstrate that SGANet achieves state-of-the-art performance in both anomaly detection and localization, confirming its effectiveness in real-world industrial scenarios.

This work addresses the vulnerability of vision models to convolutional perturbations—such as motion blur—at inference time, which poses significant security risks. The paper introduces the first provably robust certified training method against such perturbations by efficiently encoding convolutional distortions and integrating them into a formal verification framework, thereby overcoming the theoretical limitations of conventional empirical data augmentation strategies. On CIFAR-10, the proposed approach achieves over 80% robust accuracy against realistic levels of motion blur while maintaining standard accuracy comparable to baseline models. To the best of our knowledge, this is the first method to enable provably robust training against convolutional perturbations.

This study addresses the trade-off in image–point cloud registration between insufficient inliers and an excessively high outlier ratio caused by suboptimal point cloud density, which limits registration accuracy. It presents the first systematic analysis of how point cloud density affects cross-modal registration and introduces a cross-coordinate correspondence pruning mechanism. Specifically, coarse correspondences are projected into the image coordinate system, where a lightweight network fuses geometric and feature information to predict inlier confidence scores for effective outlier rejection. Furthermore, a multi-density point cloud ensemble strategy is employed to enhance inlier recall. The proposed method consistently outperforms existing approaches across multiple benchmarks, achieving a registration recall improvement of at least 8.6%.