occlusion handling

To address the significant performance degradation of object detection under occlusion, this paper proposes a robust detection framework leveraging scene context modeling. Methodologically, it introduces two interpretable information fusion mechanisms: dynamic network selection prior to prediction and scene-aware score fusion post-prediction. These are tightly coupled with a scene classification network and an RPN-DCNN backbone, enabling cross-dataset transfer without requiring large-scale scene-level annotations. The key innovation lies in explicitly incorporating scene context into both the model architecture and post-processing stages, thereby balancing structural adaptability and knowledge interpretability. Experiments on challenging partially occluded datasets demonstrate substantial improvements in both recall and precision. Moreover, joint training on mixed occluded and non-occluded images consistently outperforms single-scenario training strategies. This work provides a novel conceptual framework and a practical technical pathway toward occlusion-robust object detection.

This work addresses the limited generalization of existing generic object trackers in unseen scenarios and their coarse modeling of occlusion, which lacks fine-grained reasoning. To overcome these limitations, we propose the GOT-JEPA framework, which introduces the Joint-Embedding Predictive Architecture (JEPA) to visual tracking for the first time. Leveraging a teacher–student self-supervised mechanism, GOT-JEPA predicts pseudo-labels consistent with clean reference frames even from occluded or corrupted inputs, thereby achieving strong generalization. Furthermore, we design the OccuSolver module, which integrates point tracking with object priors to perform iterative, fine-grained visibility estimation and occlusion modeling. Extensive experiments demonstrate that our approach significantly enhances tracking robustness across seven benchmark datasets, particularly excelling in challenging occlusion and interference scenarios.

Severe occlusion leads to significant loss and corruption of target information, substantially degrading classification performance. This work proposes an occlusion-agnostic and severity-adaptive classification method that enhances robustness during training through multi-level random masking and, at test time, dynamically masks disruptive regions based on visual anomaly detection to estimate occlusion severity and adaptively select the optimal model. To the best of our knowledge, this is the first approach to jointly achieve occlusion-type invariance and severity-aware optimization. On occluded images, the proposed method improves AUC_occ by 18.5% over standard training and by 23.7% compared to fine-tuning without occlusion.

This work investigates the spatial reasoning capabilities of vision-language models (VLMs) under occlusion, specifically their ability to accurately infer the number of occluded objects. Method: We introduce CAPTURe—a novel pattern-continuation counting benchmark explicitly designed for occlusion modeling—comprising a dual-track dataset that integrates human-curated real-world images with controllable synthetic scenes, clearly distinguishing between natural and synthetic occlusion. Experiments evaluate leading VLMs including GPT-4o and Intern-VL2. Results: All models exhibit substantial performance degradation under occlusion (even GPT-4o fails), whereas human subjects achieve near-zero error. Incorporating occlusion-location priors improves accuracy, confirming that the core limitation lies in the joint failure of spatial completion and numerical reasoning. This study provides the first systematic evidence of a fundamental spatial modeling deficit in VLMs and establishes a scalable, principled evaluation paradigm for spatial cognition research.

To address key challenges in multi-object tracking under partial occlusion—namely inaccurate motion estimation, degraded appearance features, and unfair data association—this paper proposes an adaptive occlusion-aware tracking framework. Methodologically: (1) an improved Kalman filter with anomaly motion suppression is designed to enhance motion prediction robustness during occlusion; (2) a pose-keypoint-guided part-based re-identification module is introduced to mitigate appearance mismatch caused by local occlusion; and (3) an occlusion-aware IoU-appearance weighted joint association strategy is formulated to ensure fair and stable data association. Extensive experiments on MOTChallenge and DanceTrack benchmarks demonstrate significant improvements over state-of-the-art methods: IDF1 increases notably, ID switches are substantially reduced, and visualizations confirm markedly enhanced occlusion robustness.

This work addresses the limitations of generic object tracking under severe deformation, complex interference, drastic environmental changes, or unseen object categories, where insufficient generalization and online adaptation often lead to failure. To bridge this gap, the study presents a unified framework that systematically integrates key components of human visual perception—namely prior knowledge, spatial geometric constraints, and semantic context—by synergistically combining online adaptive learning, geometric modeling, and discriminative optimization. The proposed approach substantially enhances the tracker’s discriminative power, robustness, and geometric reasoning capabilities in open and dynamic environments. Extensive experiments demonstrate significant improvements in tracking accuracy across diverse challenging scenarios, effectively narrowing the performance gap between machine-based tracking and human visual perception.

Existing vision-language-action (VLA) models suffer significant performance degradation in real-world scenarios involving object occlusion, yet current evaluation benchmarks lack systematic consideration of occlusion. To address this gap, this work presents the first systematic investigation into scene-induced occlusion effects on VLA models, introducing LIBERO-Occ—a novel occlusion-oriented benchmark—and proposing the View Imagination Mechanism (VIM). VIM enhances robust action decision-making without requiring additional cameras by generating complementary viewpoints from the occluded primary view and fusing observed and imagined visual information. Extensive experiments demonstrate that VIM consistently improves model performance across diverse tasks, occlusion types, and severity levels, validating the efficacy of view imagination for perceptual completion in embodied AI.

This study addresses the widespread misuse of confidence scores from open-vocabulary detectors—such as Grounding DINO, OWLv2, and SAM3—as proxies for object visibility, when in fact these scores reflect category presence rather than the actual visibility of a specific instance. By constructing a ground-truth visibility benchmark using a geometric segmentation oracle and evaluating across multiple simulated environments and real-world video data, the work systematically audits current models and reveals, for the first time, that detector confidence remains high even when only 1/8 of an object is visible. This fundamental mismatch leads to a nearly tenfold underestimation of active perception performance. The authors argue that confidence-based evaluation and gating mechanisms are inherently biased and advocate instead for object-anchored visibility signals. To support future research, they release the first controllable occlusion benchmark dataset.

This work addresses the challenge of robust 6-degree-of-freedom object pose tracking in RGB-D data under severe occlusion and rapid motion, where existing methods often fail and struggle to recover autonomously. To this end, we propose a novel tracking framework that integrates learned keypoint matching with optimization-driven alignment. A reliability monitoring mechanism is introduced to detect tracking failure in real time and trigger a global re-detection and verification process, enabling fully automatic re-initialization without human intervention. Evaluated on both standard and newly constructed challenging datasets, our method achieves state-of-the-art accuracy at 57.6 FPS and demonstrates significantly superior robustness and sustained tracking capability in scenarios involving heavy occlusion and high-speed motion.

This work addresses the challenge of partial occlusion in 2D multi-object tracking, which often leads to ambiguity in association costs and degrades trajectory analysis and counting accuracy. To this end, the authors propose OA-SORT, a plug-and-play, training-free occlusion-aware framework that explicitly models occlusion states and dynamically adjusts association costs and motion estimation. OA-SORT introduces three novel components: an Occlusion-Aware Module (OAM), Occlusion-Aware Offset (OAO), and Bias-Aware Momentum (BAM). By incorporating Gaussian maps to suppress background interference, the framework seamlessly integrates into existing trackers. Evaluated on DanceTrack, OA-SORT achieves 63.1% HOTA and 64.2% IDF1. When integrated with four mainstream trackers, it yields average improvements of 2.08% in HOTA and 3.05% in IDF1.