Yiru Chen

Kuaishou (China)
2 previous affiliations
2Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Yiru Chen focuses on computer vision and multimodal information retrieval, with a specific interest in image processing and transformer-based video analysis. Their work spans from fundamental low-level vision tasks to high-level cross-modal representation learning, addressing the technical challenges of bridging the semantic gap between raw visual data and natural language descriptions.

Research Trajectory

The available research record indicates a transition from classical image processing techniques to modern deep learning architectures. In 2013, Chen’s work addressed the problem of background subtraction, a fundamental task in computer vision used to identify moving objects in video sequences. In the paper A novel background subtraction method based on color invariants, Chen (as second author) explored the use of photometric models to improve the robustness of foreground detection. Traditional background subtraction methods often struggle with illumination changes, such as shadows or sudden shifts in lighting, which can be incorrectly flagged as motion. By leveraging color invariants—features that remain constant under varying lighting conditions—the proposed method aimed to provide a more stable representation of the scene, thereby reducing false positives in dynamic environments.

Following a period of limited publication activity in the available corpus, Chen’s research shifted toward deep learning and multimodal information retrieval. In 2021, Chen participated in the development of the Hierarchical Transformer (HiT) framework, as detailed in HiT: Hierarchical Transformer with Momentum Contrast for Video-Text Retrieval. This work addresses the complexities of matching video content with natural language descriptions, a task complicated by the high dimensionality of video data and the semantic gap between visual and textual modalities.

The HiT approach introduces two primary innovations: hierarchical cross-modal contrastive matching and the integration of momentum contrast. The hierarchical matching mechanism allows the model to exploit the different feature characteristics present across various layers of the transformer architecture, ensuring that both low-level visual patterns and high-level semantic concepts are aligned with the text. Furthermore, the use of momentum contrast helps overcome the limitations of mini-batch sizes in end-to-end training by maintaining a consistent dictionary of negative samples, which facilitates more effective contrastive learning and improves overall retrieval performance.

Current Directions

Chen’s research contributes to the broader field of Translating Visual Interactions into Formal Graph Queries. While the specific papers focus on the algorithmic underpinnings of image processing and video retrieval, they address the fundamental challenge of the "semantic gap"—the disconnect between raw data formats and the high-level conceptual understanding required by domain experts. By improving how visual data is queried and retrieved through natural language or invariant features, this work supports the development of more effective tools for navigating complex, high-dimensional datasets.

Representative Work

Profile updated

See something wrong? Report a mistake