Hi there 👋 I am Shuang Li!

DMTA adapts modality and temporal cues dynamically for robust video-based visible-infrared person re-identification across cross-modal sequence variations.

DCASR performs difficulty-conditioned, attribute-specific restoration to adaptively enhance low-light images under diverse degradation levels.

HHA models video-based visible-infrared person re-identification with hierarchical hyperbolic geometry for robust spatio-temporal representation and modality alignment.

STHF constructs sequence-level high-frequency representations and enhances local details for robust video-based visible-infrared person re-identification.

PRL injects physical constraints from anisotropic diffusion equations into segmentation optimization for stronger generalization across segmentation tasks.

PEFC-Net enhances perceptual quality and calibrates focus regions to improve small object detection under challenging visual conditions.

HPL unifies image-to-image and text-to-image person re-identification through task-aware prompt modeling and cross-modal prompt regularization.

DSC leverages complementary motion and appearance cues to refine pseudo labels and learn robust cross-modal video ReID representations without target labels.

PortraitSR follows a progressive structure-to-texture restoration strategy for semantically consistent and visually realistic face super-resolution.

This work explores language-driven video-level representation learning for robust visible-infrared person re-identification across modality and temporal variations.

Shape-centered cues are introduced to reduce modality discrepancy and improve identity representation for visible-infrared person re-identification.

This work uses DINOv2 visual priors to learn gait representations complementary to appearance cues for video-based cross-modal retrieval.

This method improves domain adaptive person ReID through logical relation inference and multiview information interaction for reliable pseudo-label learning.

This work builds cross-compatible embeddings and semantic-consistent features to bridge sketch and image modalities for person re-identification.

This method aligns body-part representations across domains within a transformer framework for robust domain-adaptive person re-identification.