Export
YiQing Cai
3 previous affiliations
- ByteDance (China)2022–2024
- ByteDance · 2022–2024
- East China Normal University2019–2022
- MS student · 2019–2022
- Undergrad student · 2015–2019
Research Overview
YiQing Cai focuses on computer vision, specifically addressing crowd counting and scene graph generation in complex environments. Initial work explored domain adaptation for cross-domain image analysis, while subsequent research has emphasized temporal consistency in video sequences and hierarchical representations for scene understanding.
Research Trajectory
In 2021, Cai's research primarily addressed the challenges of density estimation in computer vision, specifically focusing on unsupervised cross-domain crowd counting. A significant challenge in this area is the reliance on synthetic datasets, which introduces distribution gaps when models are applied to real-world target domains. As the first author of Leveraging Intra-Domain Knowledge to Strengthen Cross-Domain Crowd Counting, Cai led the development of a two-step domain adaptation method. This work distinguished itself by addressing not only the inter-domain gap between source and target data but also the intra-domain gap within the target data itself, utilizing multi-level feature response branches to improve the model's adaptability.
The following year, Cai extended this focus on crowd counting from static images to video sequences. Recognizing that background interference and occlusion cause instability in frame-by-frame density predictions, Cai investigated methods to leverage the inherent consistency across adjacent frames. In the first-author paper Global Representation Guided Adaptive Fusion Network for Stable Video Crowd Counting, Cai proposed a network designed to establish long-term temporal representations among consecutive frames. This approach aimed to mitigate prediction mutations and improve the overall stability of video-based density estimation.
During this same period, Cai also contributed to research in complex scene understanding. As a co-author of DH-GCN: Saliency-Aware Complex Scene Graph Generation Using Dual-Hierarchy Graph Convolutional Network, Cai participated in developing a top-down approach to scene graph generation. This work addressed the poor performance of existing models in realistic, cluttered environments by utilizing salient object detectors to hierarchize objects, thereby reducing redundant information during feature optimization.
Current Directions
Based on recent thematic clusters in the evidence from 2024, Cai's focus has expanded into the architecture, training, and spatial-temporal grounding of large multimodal models. A major current direction involves the research area of Progressive Multi-Stage Training for Multimodal Alignment, which addresses the stability challenges of teaching pre-trained text models to process visual information without catastrophic forgetting. This aligns with broader work in the theme of Modeling Interleaved Image and Text Sequences, moving beyond isolated image-caption pairs to process complex, woven multimodal documents.
Additionally, Cai's recent trajectory indicates an interest in how foundation models understand space and time. This includes contributions to the research direction of Spatial Grounding via Textual Coordinate Sequences, where object localization is treated as a sequence generation problem rather than traditional regression. In the video domain, this extends to the area of Fine Grained Temporal Modeling for Video Transformers, which seeks to overcome the limitations of sparse frame sampling to capture sub-second transitions. To support the efficient training and deployment of these high-capacity models, Cai is also engaged with themes such as Scaling Model Capacity with Mixture of Low Rank Adapters and the research direction of Overcoming the Static Teacher Bottleneck in Knowledge Distillation.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake