we propose DMD, a minutiae-anchored local dense representation which captures both fine-grained ridge textures and discriminative minutiae features in a spatially structured manner. Specifically, descriptors are extracted from local patches centered and oriented on each detected minutia, forming a three-dimensional tensor, where two dimensions represent spatial locations on the fingerprint plane and the third encodes semantic features.

Emotion recognition has long grappled with the inherent subjectivity and open-ended nature of human affect, where predefined taxonomies falter against the vast, evolving spectrum of emotional expression. We present ZeroES, a Zero-Shot Ensemble framework that redefines open-vocabulary video emotion recognition by leveraging the raw capacity of large-scale vision-language models (VLMs) without task-specific optimization.

We propose a robust Mixture of Experts (MoE) emotion recognition framework that integrates diverse modalities, including Vision-Language Models and Action Unit information. Using consensus-based pseudo-labeling and a two-stage training process, we enhance label quality and reduce bias through multi-expert voting and rule-based re-ranking for human-aligned predictions.

We propose a novel framework for open-ended video question answering that enhances reasoning depth and robustness in complex real-world scenarios, as benchmarked on the CVRR-ES dataset.

We introduce a partial fingerprint pose estimation framework that leverages the collaborative potential of Dual-modal guidance from Ridge patches And Capacitive images to Optimize the feature extraction, fusion and representation. Several simple but effective strategies and mechanisms are introduced, including knowledge transfer, MoE, and decoupled probability distribution, to enhance the network's capacity for information mining and interaction.

In this work, we propose a fixed-length dense descriptor of fingerprints, and introduce FLARE—a fingerprint matching framework that integrates the Fixed-Length dense descriptor with pose-based Alignment and Robust Enhancement. This fixed-length representation employs a three-dimensional dense descriptor to effectively capture spatial relationships among fingerprint ridge structures, enabling robust and locally discriminative representations.

In this paper, we estimate the 2D finger pose using a multimodal network and map it to a standardized UV space, followed by nearly lossless mapping to 3D space using simple polynomial functions. We further highlight the applicability and appeal of finger pose in enhancing interactive experiences, and develop several prototypes to demonstrate the potential for interaction.

We demonstrate that 3D pose information of contactless fingerprints can be utilized to enhance the robustness and performance of existing recognition systems by guiding the acquisition process and constraining finger poses.

We present the runner-up solution for the Ego4D EgoSchema Challenge at CVPR 2025 (Confirmed on May 20, 2025). Inspired by the success of large models, we evaluate and leverage leading accessible multimodal large models and adapt them to video understanding tasks via few-shot learning and model ensemble strategies.

A novel framework for joint partial fingerprint identity verification and pose alignment of partial fingerprint pairs is proposed, which utilizes a multi-task CNN-Transformer hybrid network and a pre-training task on enhancement.

Latent fingerprint matching is a daunting task, primarily due to the poor quality of latent fingerprints. In this study, we propose a deep-learning based dense minutia descriptor (DMD) for latent fingerprint matching.

We propose a Phase-aggregated Dual-branch Registration Network to combine the strengths of traditional fingerprint dense registration methods and deep learning.

We proposed an end-to-end network to directly estimate a dense distortion field instead of its low dimensional representation, from a single fingerprint.

We proposed an end-to-end network to directly estimate a dense distortion field from a single fingerprint instead of its low dimensional representation.

We proposed a visualization and pose-specific unfolding method for 3D fingerprints, which can improve the compatibility between 3D and 2D fingerprints in recognition.