Wei You

Huawei Technologies (China)
3 previous affiliations
7Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Wei You focuses on computer vision, multimodal feature learning, and pattern recognition across diverse data domains. Early work addressed heterogeneous signal fusion and cross-modality retrieval, while subsequent research advanced hybrid active contour models for image segmentation. Most recently, the focus has shifted toward self-supervised feature learning and ensemble methods for 3D human action recognition.

Research Trajectory

Wei You’s research trajectory is characterized by a consistent focus on extracting and fusing complex features from challenging data modalities, evolving from early applications in bio-signal processing and smart grids to core computer vision and human action recognition.

Between 2017 and 2019, acting primarily as a senior investigator, You explored how to integrate heterogeneous data sources for specialized classification and retrieval tasks. In 2017, this involved addressing the critical issue of mental workload estimation for high-stakes professions such as pilots and surgeons. In Feature Weight Driven Interactive Mutual Information Modeling for Heterogeneous Bio-Signal Fusion to Estimate Mental Workload, You directed the development of an interactive mutual information modeling approach to fuse EEG and ECG signals, improving classification accuracy during n-back tasks. The following year, the research focus expanded into computer vision and cross-modality challenges. Specifically, You led work on large-scale face image-video retrieval, a domain complicated by the distinct mathematical properties of the media involved. In Implicit relative attribute enabled cross-modality hashing for face image-video retrieval, the proposed method successfully embedded face images (represented in Euclidean space) and face videos (represented on a Riemannian manifold) into a unified, compact Hamming space to facilitate fast retrieval. During this period, You also applied advanced pattern recognition techniques to smart grid infrastructure, serving as senior author on Non-Intrusive Load Monitoring by Voltage-Current Trajectory Enabled Transfer Learning.

By 2020, You’s focus transitioned deeper into foundational image processing, specifically addressing the theoretical limitations of active contour models used for image segmentation. Hybrid active contour models, which combine both region and edge information, often struggle with the empirical and unstable selection of weighting factors for their energy terms. This instability becomes particularly problematic during multi-object segmentation, which requires the simultaneous computation of multiple level set functions. As senior author, You tackled this theoretical gap in A New Hybrid Level Set Approach, developing a more robust framework for constructing hybrid active contours with a clearer theoretical basis for collaboration between region and edge data.

In 2022, the trajectory shifted significantly toward skeleton-based 3D human action recognition, with You taking a leading role as first author. This domain presents unique challenges due to the high diversity of human movement and the inherent noise in skeletal data representations. To overcome the reliance on extensive labeled datasets, You investigated self-supervised learning paradigms. In View Enhanced Jigsaw Puzzle for Self-Supervised Feature Learning in 3D Human Action Recognition, You designed a novel pretext task that introduces multi-view spatial information into a jigsaw puzzle format. This approach forces the feature encoder to learn view-independent, high-level representations of human skeletons, which are then integrated using a specialized view pooling encoder.

Complementing this self-supervised approach, You also addressed the limitations of relying on a single network model or a single feature type for action recognition. In Generic enhanced ensemble learning with multi-level kinematic constraints for 3D action recognition, You proposed an ensemble framework designed to maximize the extraction of valid kinematic information. By applying multi-level kinematic constraints, this work successfully mitigated the interference of irrelevant data, demonstrating a comprehensive approach to modeling complex 3D action sequences.

Current Directions

The most recent papers in the evidence, published in 2022, indicate a dedicated focus on 3D human action recognition using skeletal data. This current direction investigates how to extract robust, view-independent features from human skeleton sequences, which are often noisy and highly variable.

A primary open problem addressed in this recent research is the design of effective self-supervised pretext tasks—such as multi-view jigsaw puzzles—that can train sophisticated feature encoders without the need for exhaustive manual annotation. Alongside self-supervised feature learning, current efforts emphasize the development of advanced ensemble learning frameworks. By integrating multi-level kinematic constraints, this ongoing work seeks to synthesize multiple data modalities and deep network architectures. The overarching goal of these recent methodologies is to ensure that machine learning models can fully exploit the rich spatial and temporal dynamics embedded in human skeletal movements, overcoming the limitations of single-model and single-feature approaches.

Profile updated

See something wrong? Report a mistake