Sanping Zhou
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sanping Zhou has a total of 179 co-authors. Here are the top co-authors by number of shared publications:
- Le Wang
Xi’an Jiaotong University
45 shared publications - Gang Hua
Dolby Laboratories, Bellevue, WA, USA
26 shared publications - Nanning Zheng
Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an, China
24 shared publications - Jinjun Wang
Xi'an Jiaotong University; Xpeng Inc, China
17 shared publications - Wei Tang
Department of Computer Science, University of Illinois Chicago, Chicago, IL, USA
14 shared publications Le Wang
10 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9265-9274
Continuous Masked Diffusion for Autonomous Driving
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9265-9274
Open AccessContinuous Masked Diffusion for Autonomous Driving
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (11) , 14315-14332
Integrating Agglomerative Perception With One-Step Action Generation for Robotic Manipulation
Developing generalizable robotic policies that balance inference efficiency, manipulation accuracy, and robustness remains a formidable challenge. Existing Vision-Language-Action models demand prohibitive data scales, while keyframe-based approaches struggle to reconcile the expressivity of generative models with the latency of iterative sampling. To address this trilemma, we present Flow2Act, a unified framework that integrates agglomerative perception with a deterministic one-step generative policy. Unlike prior methods relying on separate semantic encoders or iterative diffusion processes, our approach introduces three key innovations. First, we employ an agglomerative multi-teacher visual backbone that distills complementary strengths from diverse foundation models, capturing semantics, spatial structure, and segmentation to yield robust representations without task-specific pretraining. Second, we propose a conditional MeanFlow policy that parameterizes the interval-averaged velocity field. This formulation enables genuine single-step action generation, eliminating the discretization errors and computational overhead inherent in ODE-based flow matching. Third, we devise a curriculum region-aware mechanism via a Spatial-Grounded State Space Duality architecture, which progressively shifts attention from global flow stability to fine-grained contact precision. We evaluate Flow2Act on challenging simulation benchmarks and real-world robotic tasks, demonstrating significant gains in policy performance, robustness to environmental perturbations, and cross-task real-world applicability.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (11) , 14315-14332
Integrating Agglomerative Perception With One-Step Action Generation for Robotic Manipulation
Developing generalizable robotic policies that balance inference efficiency, manipulation accuracy, and robustness remains a formidable challenge. Existing Vision-Language-Action models demand prohibitive data scales, while keyframe-based approaches struggle to reconcile the expressivity of generative models with the latency of iterative sampling. To address this trilemma, we present Flow2Act, a unified framework that integrates agglomerative perception with a deterministic one-step generative policy. Unlike prior methods relying on separate semantic encoders or iterative diffusion processes, our approach introduces three key innovations. First, we employ an agglomerative multi-teacher visual backbone that distills complementary strengths from diverse foundation models, capturing semantics, spatial structure, and segmentation to yield robust representations without task-specific pretraining. Second, we propose a conditional MeanFlow policy that parameterizes the interval-averaged velocity field. This formulation enables genuine single-step action generation, eliminating the discretization errors and computational overhead inherent in ODE-based flow matching. Third, we devise a curriculum region-aware mechanism via a Spatial-Grounded State Space Duality architecture, which progressively shifts attention from global flow stability to fine-grained contact precision. We evaluate Flow2Act on challenging simulation benchmarks and real-world robotic tasks, demonstrating significant gains in policy performance, robustness to environmental perturbations, and cross-task real-world applicability.
- 23 October 2026
- Published by Elsevier in Pattern Recognition
- Vol. 183
Towards minute-level pedestrian trajectory prediction in crowded traffic scenarios
- 23 October 2026
- Published by Elsevier in Pattern Recognition
- Vol. 183
Towards minute-level pedestrian trajectory prediction in crowded traffic scenarios
- 5 October 2026
- Published by Springer Nature
- p. 476-494
Efficient Camera Pose Augmentation for View Generalization in Robotic Policy Learning
- 5 October 2026
- Published by Springer Nature
- p. 476-494
Efficient Camera Pose Augmentation for View Generalization in Robotic Policy Learning
- 10 September 2026
- Published by Springer Nature
- p. 39-56
RAGrasp: A Retrieval-Augmented Framework with Diversity-Aware Modeling for Dexterous Grasp Generation
- 10 September 2026
- Published by Springer Nature
- p. 39-56
RAGrasp: A Retrieval-Augmented Framework with Diversity-Aware Modeling for Dexterous Grasp Generation
- Open Access
Deep Learning-Based Computer-Aided Detection and Diagnosis System for Malignant Biliary Stricture (With Video)
Background/Objectives: Malignant biliary stricture (MBS) remains difficult to diagnose accurately despite digital single-operator cholangioscopy (DSOC). We developed a deep learning (DL)-based computer-aided detection (CADe) and diagnosis (CADx) system for DSOC-based MBS assessment. Methods: This retrospective multicenter study included 149 patients from one center for model development and internal validation and 25 patients from two independent centers for external evaluation. CADe used a You Only Look Once version 11 (YOLOv11) architecture to localize irregular mucosa, abnormal vasculature, and nodular protrusions defined by the Carlos Robles-Medranda and Mendoza criteria. CADx used a Residual Network-18 classifier with gradient-weighted class activation mapping for interpretability. Results: CADe achieved a mean average precision at 50% intersection-over-union of 91.2%, with a precision of 92.0% and recall of 87.0%. CADx achieved a frame-level area under the receiver operating characteristic curve (AUC) of 0.960 in internal validation and 0.843 in external validation. External frame-level sensitivity was 52.0% and specificity was 95.2%. For the patient-level external endpoint, sensitivity was 85.7%, specificity was 94.4%, accuracy was 92.0%, and AUC was 0.881. Conclusions: The two-stage system combines localization of predefined cholangioscopic features with interpretable diagnostic classification. The small external cohort and marked reduction in frame-level sensitivity preclude firm conclusions regarding generalizability; prospective multicenter and live-procedure evaluation is required.
Open AccessDeep Learning-Based Computer-Aided Detection and Diagnosis System for Malignant Biliary Stricture (With Video)
Background/Objectives: Malignant biliary stricture (MBS) remains difficult to diagnose accurately despite digital single-operator cholangioscopy (DSOC). We developed a deep learning (DL)-based computer-aided detection (CADe) and diagnosis (CADx) system for DSOC-based MBS assessment. Methods: This retrospective multicenter study included 149 patients from one center for model development and internal validation and 25 patients from two independent centers for external evaluation. CADe used a You Only Look Once version 11 (YOLOv11) architecture to localize irregular mucosa, abnormal vasculature, and nodular protrusions defined by the Carlos Robles-Medranda and Mendoza criteria. CADx used a Residual Network-18 classifier with gradient-weighted class activation mapping for interpretability. Results: CADe achieved a mean average precision at 50% intersection-over-union of 91.2%, with a precision of 92.0% and recall of 87.0%. CADx achieved a frame-level area under the receiver operating characteristic curve (AUC) of 0.960 in internal validation and 0.843 in external validation. External frame-level sensitivity was 52.0% and specificity was 95.2%. For the patient-level external endpoint, sensitivity was 85.7%, specificity was 94.4%, accuracy was 92.0%, and AUC was 0.881. Conclusions: The two-stage system combines localization of predefined cholangioscopic features with interpretable diagnostic classification. The small external cohort and marked reduction in frame-level sensitivity preclude firm conclusions regarding generalizability; prospective multicenter and live-procedure evaluation is required.
- 18 July 2026
- Published by Elsevier in Pattern Recognition
- Vol. 180
Time-Unified Diffusion Policy with action discrimination for robotic manipulation
- 18 July 2026
- Published by Elsevier in Pattern Recognition
- Vol. 180
Time-Unified Diffusion Policy with action discrimination for robotic manipulation
- 12 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
Structured Graph Reasoning for Traffic Anomaly Detection
Unsupervised traffic anomaly detection aims to detect abnormal frames and objects in dashcam videos given only normal videos during training. Most prior works focus on reconstruction or prediction pipelines, while recent graph-based extensions still rely on graph feature learning or limited statistical interactions. Consequently, reasoning over structured relations and abnormal interactions remains largely unexplored. This paper introduces a structured graph reasoning method that bridges normality learning, anomaly detection, and structured reasoning in a unified framework. Unlike prior works, our approach interprets the graph as a semantic abstraction tailored for comprehensive traffic scene understanding. It includes three innovative components: (1) A domain-specific graph representation that models semantic relationships for road scenarios, providing a holistic description of traffic activities. (2) A graph reasoning detector that detects traffic anomalies by reasoning over structured traffic scene graphs. It integrates node anomaly, i.e., unusual object movements, and edge anomaly, i.e., abnormal object interactions, within a unified graph optimization approach, enabling the effective detection of complex, interaction-driven anomalies. (3) An Anomaly Transformer that predicts future traffic scene graphs by attending node and edge queries to object locations and relation histories, thereby effectively learning individual and interaction-level normal traffic behaviors. Experiments on two public datasets, i.e., DoTA and MM-AU, demonstrate the superior performance of our method and its effectiveness in detecting complex abnormal traffic events.
- 12 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
Structured Graph Reasoning for Traffic Anomaly Detection
Unsupervised traffic anomaly detection aims to detect abnormal frames and objects in dashcam videos given only normal videos during training. Most prior works focus on reconstruction or prediction pipelines, while recent graph-based extensions still rely on graph feature learning or limited statistical interactions. Consequently, reasoning over structured relations and abnormal interactions remains largely unexplored. This paper introduces a structured graph reasoning method that bridges normality learning, anomaly detection, and structured reasoning in a unified framework. Unlike prior works, our approach interprets the graph as a semantic abstraction tailored for comprehensive traffic scene understanding. It includes three innovative components: (1) A domain-specific graph representation that models semantic relationships for road scenarios, providing a holistic description of traffic activities. (2) A graph reasoning detector that detects traffic anomalies by reasoning over structured traffic scene graphs. It integrates node anomaly, i.e., unusual object movements, and edge anomaly, i.e., abnormal object interactions, within a unified graph optimization approach, enabling the effective detection of complex, interaction-driven anomalies. (3) An Anomaly Transformer that predicts future traffic scene graphs by attending node and edge queries to object locations and relation histories, thereby effectively learning individual and interaction-level normal traffic behaviors. Experiments on two public datasets, i.e., DoTA and MM-AU, demonstrate the superior performance of our method and its effectiveness in detecting complex abnormal traffic events.
- 18 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
DiffLabel: Robust Pseudo Label Generation for Semi-Supervised Temporal Language Grounding
Semi-Supervised Temporal Language Grounding (SS-TLG) aims at retrieving a target moment from untrimmed videos by a given language query with large-scale unlabeled videos. Similar to most semi-supervised learning tasks, the main challenge in SS-TLG is how to improve the quality of pseudo labels. Existing methods are affected by temporal location biases from datasets, which leads to unreliable pseudo labels, reducing the performance of whole semi-supervised learning process. In this paper, we propose DiffLabel, a novel approach of utilizing the diffusion model to generate robust pseudo labels for SS-TLG. Specifically, the diffusion model is incorporated into a teacher-student framework. The teacher model refines a set of randomly initialized noisy moments to the output pseudo labels in a gradual way. In the student model, target moments diffuse from pseudo labels and ground truths to random distributions, and the model learns to reverse this noising process. Due to the introduction of random noise, the model can better mitigate the interference of temporal location biases, thereby improving the robustness of pseudo labels and optimize the entire semi-supervised learning process. In addition, we propose two query-aware data perturbations to avoid semantic information redundancy and annotation deviation. Extensive experiments validate the effectiveness and generalization of our approach, which outperforms state-of-the-art methods on both Charades-CD-OOD and Activitynet-CD-OOD datasets. The code will be available.
- 18 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
DiffLabel: Robust Pseudo Label Generation for Semi-Supervised Temporal Language Grounding
Semi-Supervised Temporal Language Grounding (SS-TLG) aims at retrieving a target moment from untrimmed videos by a given language query with large-scale unlabeled videos. Similar to most semi-supervised learning tasks, the main challenge in SS-TLG is how to improve the quality of pseudo labels. Existing methods are affected by temporal location biases from datasets, which leads to unreliable pseudo labels, reducing the performance of whole semi-supervised learning process. In this paper, we propose DiffLabel, a novel approach of utilizing the diffusion model to generate robust pseudo labels for SS-TLG. Specifically, the diffusion model is incorporated into a teacher-student framework. The teacher model refines a set of randomly initialized noisy moments to the output pseudo labels in a gradual way. In the student model, target moments diffuse from pseudo labels and ground truths to random distributions, and the model learns to reverse this noising process. Due to the introduction of random noise, the model can better mitigate the interference of temporal location biases, thereby improving the robustness of pseudo labels and optimize the entire semi-supervised learning process. In addition, we propose two query-aware data perturbations to avoid semantic information redundancy and annotation deviation. Extensive experiments validate the effectiveness and generalization of our approach, which outperforms state-of-the-art methods on both Charades-CD-OOD and Activitynet-CD-OOD datasets. The code will be available.
- 2 May 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Action hints: Semantic typicality and context uniqueness for generalizable skeleton-based video anomaly detection
- 2 May 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Action hints: Semantic typicality and context uniqueness for generalizable skeleton-based video anomaly detection