Selected work

Representative Papers

Generative Semantic Segmentation via an Observable Semantic-Image Interface and Hierarchical Generator Evidence Alignment

Aug 11, 2026

This work addresses color drift, boundary ambiguity, and the limitation of latent-feature decoders treating images merely as intermediate visualizations in generative semantic segmentation. To overcome these issues, the authors propose the Semantic Prism framework, which employs diffusion distillation to construct a single-step generator, integrates a fixed class-color codebook to establish an explicit probabilistic interface, and introduces a hierarchical evidence alignment mechanism in logit space to predict residuals while preserving the image-defined interface as the reference for the final distribution. The study also innovatively presents C-IHD, an error-ranking method that requires no additional predictor. Experiments demonstrate that the approach achieves 72.07% mIoU (+11.39) with an ECE of 0.41% on Cityscapes, and attains 62.22% and 46.89% mIoU on BDD100K and ACDC, respectively, while significantly improving pixel-wise error ranking AUPR—e.g., from 0.6580 to 0.7557 on ACDC.

0 citationsRead paper

Cross-View Sequential Visual Localization with Spatio-Temporal Context Modeling for Autonomous Driving

Aug 11, 2026

This work addresses the limitations of existing cross-view visual localization methods, which often neglect temporal information and consequently suffer degraded performance under dynamic occlusions, illumination variations, and repetitive textures. To overcome this, we propose the first sequential localization framework that integrates spatiotemporal context by recursively aggregating historical states through a cross-frame module to enhance ground-level features in the current frame. Coupled with hierarchical fine-grained feature representations, our approach enables robust candidate region classification and precise offset estimation from satellite imagery. By introducing temporal modeling into cross-view localization for the first time, our method significantly improves both robustness and accuracy over single-frame baselines. Experiments demonstrate a reduction in average localization error from 3.80 m to 1.57 m on the CVIS dataset, with R@1m improving from 8.14% to 40.22%. Moreover, it achieves zero-shot transfer to KITTI-CVL with a 2.61 m error and yields a real-world vehicle test error of 2.84 m, attaining R@5m of 96.86%.

0 citationsRead paper

Structure-Enhanced Features and Quality-Aware Dynamic Anchor Scoring for Robust Lane Detection

Aug 10, 2026

This work addresses structural discontinuities in lane detection caused by occlusions and complex scenes, as well as the misalignment between classification confidence and localization quality that leads to false positives and missed detections. Without altering the inference pipeline of the Anchor Decomposition Network (ADNet), the authors propose two key enhancements: first, a Gated Horizontal-Vertical Token module is introduced to strengthen directional structural feature continuity in the backbone network; second, a Line-Quality-Aware dynamic anchor scoring mechanism is devised to recalibrate anchors based on quality supervision, hard negative suppression, and pairwise ranking—eliminating the need for additional network branches. Evaluated on VIL-100, the method improves ADNet-R34’s F1@50 from 89.97 to 91.28, while experiments on CULane and TuSimple further confirm its effectiveness and low computational overhead.

0 citationsRead paper

PromptShield Home: Ambient Multimodal Prompt Injection Defense for Smart-Home Agents

Aug 05, 2026

This work addresses the vulnerability of smart home agents to multimodal prompt injection attacks—such as voice from televisions, on-screen text, or overheard conversations—that can trigger unintended actions. It presents the first systematic characterization of such threats in domestic environments and introduces PromptShield-Home, a benchmark encompassing representative scenarios including ambiguous addressing, audiovisual injection, and spurious health-monitoring triggers. The study evaluates defense efficacy across three abstraction levels: conventional detectors (L0), monolithic multimodal large models (L1), and multi-agent collaborative mechanisms (L2). Experiments reveal that even the best single-method approach achieves only a 76.5% safe execution rate, whereas an ideal routing strategy raises this to 94.1%. Notably, detectors tend toward over-execution while multimodal models frequently reject legitimate commands, exhibiting complementary failure modes that underscore the necessity of integrating learned routing with multi-sensor collaboration.

0 citationsRead paper

COMEX: A Composition-Grounded Benchmark and Learning Framework for Explainable Aesthetic Image Cropping

Aug 03, 2026

This work addresses the common oversight of compositional aesthetics in existing image cropping methods, which often fail to produce results that are both visually appealing and interpretable. To overcome this limitation, the authors propose a novel composition-driven paradigm for explainable image cropping, reframing the task as a joint modeling of cropping, composition classification, and natural language explanation generation. They introduce COMEX, the first large-scale benchmark annotated with quadruplets comprising images, crop bounding boxes, composition categories, and human-written explanations. The dataset is constructed via image expansion and IO inversion strategies, and a two-stage training framework—combining supervised fine-tuning (SFT) and group relative policy optimization (GRPO)—is employed to optimize multi-task outputs. Experiments demonstrate that the proposed approach significantly outperforms existing methods in cropping quality, composition prediction accuracy, and explanation faithfulness, while also providing a unified evaluation platform for 15 vision-language models.

0 citationsRead paper

Recent publications

Latest Papers

Generative Semantic Segmentation via an Observable Semantic-Image Interface and Hierarchical Generator Evidence Alignment

Aug 11, 2026

This work addresses color drift, boundary ambiguity, and the limitation of latent-feature decoders treating images merely as intermediate visualizations in generative semantic segmentation. To overcome these issues, the authors propose the Semantic Prism framework, which employs diffusion distillation to construct a single-step generator, integrates a fixed class-color codebook to establish an explicit probabilistic interface, and introduces a hierarchical evidence alignment mechanism in logit space to predict residuals while preserving the image-defined interface as the reference for the final distribution. The study also innovatively presents C-IHD, an error-ranking method that requires no additional predictor. Experiments demonstrate that the approach achieves 72.07% mIoU (+11.39) with an ECE of 0.41% on Cityscapes, and attains 62.22% and 46.89% mIoU on BDD100K and ACDC, respectively, while significantly improving pixel-wise error ranking AUPR—e.g., from 0.6580 to 0.7557 on ACDC.

0 citationsRead paper

Cross-View Sequential Visual Localization with Spatio-Temporal Context Modeling for Autonomous Driving

Aug 11, 2026

This work addresses the limitations of existing cross-view visual localization methods, which often neglect temporal information and consequently suffer degraded performance under dynamic occlusions, illumination variations, and repetitive textures. To overcome this, we propose the first sequential localization framework that integrates spatiotemporal context by recursively aggregating historical states through a cross-frame module to enhance ground-level features in the current frame. Coupled with hierarchical fine-grained feature representations, our approach enables robust candidate region classification and precise offset estimation from satellite imagery. By introducing temporal modeling into cross-view localization for the first time, our method significantly improves both robustness and accuracy over single-frame baselines. Experiments demonstrate a reduction in average localization error from 3.80 m to 1.57 m on the CVIS dataset, with R@1m improving from 8.14% to 40.22%. Moreover, it achieves zero-shot transfer to KITTI-CVL with a 2.61 m error and yields a real-world vehicle test error of 2.84 m, attaining R@5m of 96.86%.

0 citationsRead paper

Structure-Enhanced Features and Quality-Aware Dynamic Anchor Scoring for Robust Lane Detection

Aug 10, 2026

This work addresses structural discontinuities in lane detection caused by occlusions and complex scenes, as well as the misalignment between classification confidence and localization quality that leads to false positives and missed detections. Without altering the inference pipeline of the Anchor Decomposition Network (ADNet), the authors propose two key enhancements: first, a Gated Horizontal-Vertical Token module is introduced to strengthen directional structural feature continuity in the backbone network; second, a Line-Quality-Aware dynamic anchor scoring mechanism is devised to recalibrate anchors based on quality supervision, hard negative suppression, and pairwise ranking—eliminating the need for additional network branches. Evaluated on VIL-100, the method improves ADNet-R34’s F1@50 from 89.97 to 91.28, while experiments on CULane and TuSimple further confirm its effectiveness and low computational overhead.

0 citationsRead paper

PromptShield Home: Ambient Multimodal Prompt Injection Defense for Smart-Home Agents

Aug 05, 2026

This work addresses the vulnerability of smart home agents to multimodal prompt injection attacks—such as voice from televisions, on-screen text, or overheard conversations—that can trigger unintended actions. It presents the first systematic characterization of such threats in domestic environments and introduces PromptShield-Home, a benchmark encompassing representative scenarios including ambiguous addressing, audiovisual injection, and spurious health-monitoring triggers. The study evaluates defense efficacy across three abstraction levels: conventional detectors (L0), monolithic multimodal large models (L1), and multi-agent collaborative mechanisms (L2). Experiments reveal that even the best single-method approach achieves only a 76.5% safe execution rate, whereas an ideal routing strategy raises this to 94.1%. Notably, detectors tend toward over-execution while multimodal models frequently reject legitimate commands, exhibiting complementary failure modes that underscore the necessity of integrating learned routing with multi-sensor collaboration.

0 citationsRead paper

COMEX: A Composition-Grounded Benchmark and Learning Framework for Explainable Aesthetic Image Cropping

Aug 03, 2026

This work addresses the common oversight of compositional aesthetics in existing image cropping methods, which often fail to produce results that are both visually appealing and interpretable. To overcome this limitation, the authors propose a novel composition-driven paradigm for explainable image cropping, reframing the task as a joint modeling of cropping, composition classification, and natural language explanation generation. They introduce COMEX, the first large-scale benchmark annotated with quadruplets comprising images, crop bounding boxes, composition categories, and human-written explanations. The dataset is constructed via image expansion and IO inversion strategies, and a two-stage training framework—combining supervised fine-tuning (SFT) and group relative policy optimization (GRPO)—is employed to optimize multi-task outputs. Experiments demonstrate that the proposed approach significantly outperforms existing methods in cropping quality, composition prediction accuracy, and explanation faithfulness, while also providing a unified evaluation platform for 15 vision-language models.

0 citationsRead paper