Export
Errui Ding
· 4 previous affiliations
- China Mobile Research InstituteSince 2025
- Researcher · Since 2025
- Baidu (China)2011–2025
- Director · 2011–2025
- ArcSoft2008–2011
- Researcher · 2008–2011
- Xidian University1998–2008
- PhD student · 2002–2008
- Undergrad student · 1998–2002
Research Overview
Errui Ding focuses on computer vision, multi-modal understanding, and image generation. His research trajectory spans foundational work in temporal video localization and document intelligence to advanced architectures for 3D object detection and vision transformers. Most recently, his work investigates unified vision models guided by explanatory instructions and high-fidelity generative synthesis for complex human-object interactions.
Research Trajectory
Errui Ding’s early research focused on identifying discriminative features in complex visual and temporal data. With collaborators, he explored attribute-guided attention for fine-grained recognition in Localizing by Describing: Attribute-Guided Attention Localization for Fine-Grained Recognition (2017), contributing to the broader area of Learning Object Localization from Image Level Labels. This work utilized part attribute descriptions to train region localizers via reinforcement learning. He subsequently participated in advancing video understanding with BMN: Boundary-Matching Network for Temporal Action Proposal Generation (2019), which introduced a 2D confidence map to evaluate all possible action proposals simultaneously, significantly improving inference speed.
By 2021, his focus expanded to document intelligence and optical character recognition. He participated in the development of PGNet: Real-time Arbitrarily-Shaped Text Spotting with Point Gathering Network, a single-shot framework that bypassed computationally expensive region-of-interest operations to achieve high-speed text spotting for curved text. He also directed the development of StrucTexT: Structured Text Understanding with Multi-Modal Transformers, a unified framework for entity labeling and linking in visually rich documents, advancing the research direction of End-to-End Scene Text Spotting. Concurrently, he contributed to foundational vision architectures with MixFormer: Mixing Features across Windows and Dimensions (2022), which integrated local-window self-attention and depth-wise convolution in parallel to expand receptive fields, contributing to the area of Global Communication for Sparse Window Attention.
Between 2022 and 2023, Ding led significant efforts in 3D perception, particularly for autonomous driving applications. He oversaw the creation of Rope3D: TheRoadside Perception Dataset for Autonomous Driving and Monocular 3D Object Detection Task, a large-scale dataset captured from elevated roadside cameras designed to eliminate blind spots, driving progress in the area of Overcoming Sensor Occlusion Through Collaborative Perception. To bridge the performance gap between camera and LiDAR systems, he collaborated on StereoDistill: Pick the Cream from LiDAR for Distilling Stereo-Based 3D Object Detection, a cross-modal knowledge distillation framework that aligns with the research direction of Cross-Modal Distillation for 3D Perception. He further refined monocular 3D detection by decoupling the problem into localization and appearance sub-tasks in The Devil is in the Task: Exploiting Reciprocal Appearance-Localization Features for Monocular 3D Object Detection.
During this period, Ding also continued to advance video and image processing methodologies. He directed work on the Temporal Segment Transformer for Action Segmentation (2023), which utilized a predict-and-refine architecture to smooth noisy action boundaries. This was complemented by Learning Multi-Granular Spatio-Temporal Graph Network for Skeleton-based Action Recognition, which processed skeleton data at multiple temporal resolutions. In the image domain, he led the development of Effective Invertible Arbitrary Image Rescaling, utilizing invertible neural networks for high-quality bidirectional image rescaling.
His work on object detection architectures matured through contributions to Detection Transformers (DETR). He participated in Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment (2023), which accelerated convergence by employing multiple groups of object queries, advancing the area of Hybrid Label Assignment for Detection Transformers. Building on this, he directed KD-DETR: Knowledge Distillation for Detection Transformer with Consistent Distillation Points Sampling (2024), introducing shared object queries between teacher and student models to resolve spatial and semantic inconsistencies during compression.
Most recently, Ding has focused on generative models and unified vision frameworks. He oversaw HD-Fusion: Detailed Text-to-3D Generation Leveraging Multiple Noise Estimation (2024), which introduced a memory-efficient tiled noise estimation technique for high-resolution 3D asset generation. He also directed Explanatory Instructions: Towards Unified Vision Tasks Understanding and Zero-shot Generalization (2025), leveraging detailed linguistic descriptions of task objectives to enable models to generalize to unseen visual transformations. Alongside this, he led A Mutual Learning Method for Salient Object Detection with intertwined Multi-Supervision-Revised, a multi-task framework for real-time saliency mapping. His most recent collaborative work includes Re-HOLD: Video Hand Object Interaction Reenactment via adaptive Layout-instructed Diffusion Model (2026), which synthesizes realistic human-object interaction videos using adaptive layout guidance.
Current Directions
The most recent papers in the evidence indicate a strong focus on unified multimodal intelligence and high-fidelity generative modeling. Recent work investigates how linguistic "explanatory instructions" can replace rigid task-specific labels, allowing vision models to understand the underlying logic of visual transformations and achieve zero-shot generalization across diverse tasks. Concurrently, Ding's research addresses the complexities of generative video and 3D synthesis. This includes developing memory-efficient diffusion techniques for high-resolution text-to-3D generation and designing adaptive layout-instructed diffusion models to synthesize realistic, temporally consistent human-object interactions in video. Finally, ongoing work in salient object detection explores mutual learning frameworks that intertwine foreground and edge supervision to maintain real-time performance while improving boundary precision.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake