Export
Jingtuo Liu
1 previous affiliation
- Baidu (China)2011–2025
- Chief Architect · 2011–2025
Research Overview
Jingtuo Liu focuses on computer vision and deep learning, with a primary emphasis on facial analysis, object detection, and generative modeling. His work spans the development of robust detection frameworks for unconstrained environments, the optimization of large-scale representation learning for face recognition, and the creation of high-fidelity generative systems for face swapping and lip synchronization. Liu’s research trajectory reflects a transition from discriminative tasks—such as identifying small or occluded faces—toward the synthesis of realistic facial dynamics and the acceleration of transformer-based architectures for edge deployment.
Research Trajectory
The early phase of Liu’s research, beginning around 2018, was centered on the challenges of object detection in complex environments. A significant portion of this work addressed the limitations of standard detectors when faced with small, blurred, or partially occluded subjects. During this period, Liu explored the integration of contextual information to improve detection accuracy, moving beyond local pixel data to leverage surrounding semantic cues. This work frequently utilized Multi-Scale Fusion for Preserving Fine Grained Spatial Details, a technique essential for maintaining high-resolution features across deep neural networks.
In 2019, Liu’s focus expanded to include visual text extraction and robust reading. He participated in large-scale benchmarking efforts for detecting and recognizing arbitrary-shaped text in scene images, contributing to methods that handle irregular geometries and complex backgrounds. This period also saw an engagement with Dynamic Label Assignment for Object Detection, as researchers sought more flexible ways to match ground-truth objects to neural network anchors than traditional fixed geometric rules allowed.
By 2020 and 2021, Liu’s work shifted toward the problems of scale and security in facial recognition. He contributed to research addressing face anti-spoofing, reframing the problem as anomaly detection to improve generalization against unknown spoofing types. Simultaneously, he worked on the computational and statistical challenges of training models on massive, "in-the-wild" datasets. This involved addressing The Label Noise Ceiling in Benchmark Evaluation, where the quality of ground-truth labels becomes the limiting factor for model performance, and employing Deep Metric Learning with Triplet Loss to learn discriminative embeddings across millions of identities.
Since 2022, Liu has increasingly focused on generative facial synthesis and model efficiency. This recent work includes the development of frameworks for video face swapping and audio-driven lip synchronization. These projects often utilize style-based generators to achieve high-fidelity results while maintaining identity consistency. To ensure these complex models remain viable for real-world applications, Liu has also investigated the acceleration of Vision Transformers (ViTs), focusing on heterogeneous attention patterns to reduce computational overhead without sacrificing accuracy. This evolution demonstrates a move toward The Irreversibility of Errors in Coarse-to-Fine Refinement, where generative models must carefully balance initial structural "sketches" with fine-grained detail synthesis.
Representative Work
Liu served as the senior author on PyramidBox: A Context-Assisted Single Shot Face Detector, which addressed the difficulty of detecting faces in uncontrolled environments. The research introduced "PyramidAnchors," a semi-supervised method designed to supervise high-level contextual feature learning. By combining a Low-level Feature Pyramid Network with context-sensitive anchors, the framework improved the utilization of environmental cues to identify faces that are otherwise difficult to detect due to scale or occlusion.
In the domain of document intelligence, Liu contributed to the development of EATEN: Entity-Aware Attention for Single Shot Visual Text Extraction. This work proposed an end-to-end trainable system for extracting specific "Text of Interest" (ToI) from images, such as those found on ID cards or receipts. Unlike traditional pipelines that separate optical character recognition (OCR) from structural information extraction, EATEN uses entity-aware decoders and a state transition mechanism to extract information in a single shot, enhancing the robustness of visual entity recognition.
Addressing the constraints of large-scale training, Liu participated in the development of the Dynamic Class Queue for Large Scale Face Recognition In the Wild. This research targeted the linear scaling of memory and computational costs relative to the number of identities in a dataset. By implementing a dynamic class queue (DCQ), the authors sought to mitigate the impact of long-tailed class distributions and hardware limitations, allowing for more efficient representation learning in massive facial datasets.
In the field of generative synthesis, Liu contributed to StyleSwap: Style-Based Generator Empowers Robust Face Swapping. This work leveraged the architecture of style-based generators, specifically StyleGAN2, to improve the fidelity of person-agnostic face swapping. The framework focuses on balancing information between source and target faces to reduce visible artifacts and optimize identity similarity, demonstrating that minimal modifications to existing style-based architectures can yield significant improvements in high-fidelity facial manipulation.
Liu also contributed to StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-based Generator, which addresses the challenge of syncing lip movements with arbitrary audio. The research introduced a mask-guided spatial information encoding module to preserve the details of the original face while accurately reflecting mouth shapes driven by audio input. This approach enables high-fidelity synchronization in both one-shot and few-shot scenarios, overcoming the limitations of previous models that often produced generic movement patterns.
To address the deployment of modern architectures, Liu participated in research on Accelerating Vision Transformers Based on Heterogeneous Attention Patterns. This work observed that attention patterns in ViTs vary significantly across layers, with earlier layers often exhibiting more static, image-independent patterns. The proposed "dynamic-guided static self-attention" (DGSSA) method replaces complex dynamic self-attention with more efficient static matrices in early layers, providing a compression pipeline that accelerates inference while maintaining the representative capacity of the transformer model.
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