instance segmentation

This work addresses the challenge of instance segmentation in low-texture, uniformly colored industrial scenes, where existing methods struggle due to insufficient texture and contextual cues, and where instance definitions often vary by task. The authors propose a boundary-aware few-shot instance segmentation framework that leverages a pretrained foundation model to extract visual features and employs a lightweight signed distance function (SDF) head to predict boundary-sensitive distance maps. Instance masks are then reconstructed from these SDF predictions. By shifting supervision from internal appearance to explicit boundary modeling, the approach enables flexible instance definition—such as whole objects or subparts—via mask conditioning. A pixel-wise shallow MLP facilitates rapid training and fine-grained control. Experiments demonstrate superior few-shot generalization, robustness, and precise control over segmentation granularity on low-texture industrial components and food imagery.

To address the challenges of instance segmentation in dense visual scenes—particularly those arising from occlusion, cluttered backgrounds, and scale variation—this paper proposes PerSense++, a training-free, end-to-end personalized instance segmentation framework. Methodologically, it introduces (1) density-map-guided candidate point generation coupled with adaptive spatial gating for refined proposal selection; (2) a feedback-optimized density map generation module to enhance localization accuracy; and (3) diversity-aware exemplar selection, hybrid candidate generation, and irrelevant mask rejection to improve occlusion robustness. Extensive evaluations demonstrate that PerSense++ significantly outperforms state-of-the-art methods across multiple benchmarks. Furthermore, the authors introduce PerSense-D—the first dedicated benchmark dataset for dense personalized segmentation—to foster research in this emerging direction.

To address inaccurate personalized instance segmentation in dense images caused by severe occlusion, scale variation, and cluttered backgrounds, this paper proposes the first end-to-end, training-free, model-agnostic one-shot segmentation framework. Our method comprises two core components: (1) a density-map-based Instance Detection Module (IDM) for coarse-grained localization; and (2) a Point Prompt Selection Module (PPSM) coupled with a feedback-driven exemplar self-selection mechanism, which jointly generate highly discriminative instance-level point prompts and iteratively refine the density map. Evaluated on our newly constructed benchmark PerSense-D, the framework achieves 71.61% mIoU—surpassing the state-of-the-art by up to 47.16%. This substantial improvement demonstrates significantly enhanced robustness in challenging, unconstrained real-world scenarios.

This work addresses the limitations of existing controllable image generation methods, which often conflate instance attributes in complex multi-instance scenes and rely heavily on manual annotations. To overcome these challenges, we propose InstanceControl—the first approach to enable precise multi-instance control without requiring instance-level labels. Our method leverages a vision-language model (VLM) to automatically establish instance-level correspondences between textual descriptions and visual conditions, and introduces an adaptive mask refinement mechanism that dynamically optimizes instance masks during generation. Extensive experiments demonstrate that InstanceControl significantly outperforms current state-of-the-art techniques, achieving superior instance-level control accuracy while maintaining high-fidelity image synthesis.

To address the heavy reliance of instance segmentation on costly, labor-intensive manual annotations, this paper proposes a fully unsupervised instance segmentation framework. Methodologically, it introduces the first integration of superpixels—generated via MultiCut and low-level features—with self-supervised visual representations, and designs a superpixel-guided mask loss with dual hard and soft branches. Furthermore, an adaptive-weighted self-training mechanism is incorporated to enable pseudo-label quality-driven iterative optimization. The core contributions are: (1) joint modeling of superpixels and self-supervised features; (2) a two-stage learnable mask loss function; and (3) an adaptive self-training strategy. Evaluated on standard benchmarks, the proposed method achieves state-of-the-art performance in both unsupervised instance segmentation and unsupervised object detection, outperforming all existing approaches.

This work proposes L2G-Det, a novel framework for detecting and segmenting novel object instances in open-world scenarios using only a few template images, addressing challenges such as occlusion and background clutter. Departing from conventional proposal-based paradigms, L2G-Det introduces an end-to-end local-to-global pipeline: it generates candidate points through dense local patch matching between template and query images, then leverages these refined candidates to guide an enhanced Segment Anything Model (SAM) in reconstructing complete instance masks. Key innovations include the introduction of instance-level object tokens to strengthen SAM’s instance awareness and a tailored instance-specific prompting mechanism. Experiments demonstrate that L2G-Det significantly outperforms existing proposal-based methods in complex open-world settings, achieving markedly improved robustness and accuracy in novel instance detection and segmentation.

Existing approaches struggle to balance zero-shot segmentation accuracy with interactive image matting, lacking a unified framework capable of simultaneously achieving high-quality segmentation and fine-grained alpha matte generation. This work proposes SAMA, a lightweight unified model that extends the Segment Anything Model (SAM) by incorporating a Multi-View Local Encoder (MVLE), a Localization Adapter (Local-Adapter), and a dual-task prediction head, thereby integrating interactive segmentation and matting within a single architecture for the first time. Through a joint training strategy, SAMA significantly enhances boundary detail recovery with only a marginal increase in parameters, achieving state-of-the-art performance across multiple segmentation and matting benchmarks and demonstrating its efficiency and versatility for diverse downstream tasks.

This work addresses the challenge of inaccurate forgery localization in cross-domain scenarios, particularly under text-guided high-fidelity image inpainting and arbitrary input resolutions. We propose EDGER, a patch-based dual-branch framework: an edge-guided segmentation branch enhances high-frequency inconsistencies at manipulation boundaries through frequency-domain edge detection, fusing RGB and edge features to produce pixel-level masks; a synthesis heatmap branch leverages CLIP-ViT with LoRA adapters to identify fully synthesized patches, yielding coarse-grained priors. The two branches operate synergistically to achieve resolution-agnostic, high-precision forgery localization. By jointly exploiting frequency-domain edge cues and patch-level synthesis priors—a first in the field—EDGER significantly improves cross-domain generalization and demonstrates strong performance on the MediaEval 2025 SynthIM Challenge, scaling effectively to multi-megapixel images.

Existing superpixel methods often sacrifice region regularity when leveraging deep learning for improved segmentation accuracy, thereby compromising interpretability and downstream applicability. This paper proposes SPAM, a framework that achieves high-accuracy, semantics-aware superpixel segmentation while preserving geometric regularity. Its key contributions are: (1) the first integration of large-scale, semantics-agnostic pre-trained models into superpixel generation, effectively decoupling semantic representation from structural modeling; (2) a joint semantic-structural optimization strategy that synergistically combines deep feature extraction with classical regularization to enhance boundary consistency; and (3) flexible support for arbitrary prior inputs, interactive object focusing, and adaptive refinement of uncertain regions. SPAM significantly outperforms state-of-the-art methods across multiple benchmarks, with both quantitative metrics and qualitative visualizations confirming its effectiveness. The code and models are publicly available.

This work addresses the lack of reliable uncertainty quantification in existing instance segmentation methods, which often fail to guarantee the proximity between predicted and ground-truth masks. To this end, it introduces conformal prediction—a first for instance segmentation—and proposes an algorithm with provable coverage guarantees. The method constructs, for any pixel coordinate, a confidence set of predictions that is theoretically guaranteed to contain at least one mask whose Intersection over Union (IoU) with the true instance mask exceeds a pre-specified threshold. By integrating both asymptotic and finite-sample theory, the approach achieves target coverage across diverse applications, including agricultural field delineation, cell segmentation, and vehicle detection. Moreover, it adaptively adjusts the size of the prediction set based on query difficulty and outperforms baseline methods such as Learn Then Test, conformal risk control, and morphological dilation.