Zhang Chenhao

I am a PhD student in the Visual Media Computing Lab (VMCL) within the School of Computer Science at Beijing Institute of Technology. Under the supervision of Professor Zhang Lei, my research focuses on advancing techniques in 3D vision and its applications.

Email  /  GitHub  /  Google Scholar

Research

Decoupling Multi-view Inconsistency Detection from Gaussian Splatting via 3D Consistency for Transient-Free Reconstruction

This work introduces a 3D consistency-guided framework that decouples transient detection from Gaussian Splatting optimization. By constructing a static 3D reference and deriving reliable transient masks from multi-view consistency, the method suppresses transient disturbances while preserving complete static content, enabling high-quality reconstruction from challenging unconstrained captures.

P2G: Rotation-Robust Gaussian Splatting for Panoramic-to-General Motion Novel View Synthesis in Indoor Scenes

This work introduces P2G-GS for high-quality novel view synthesis from panoramic-style inputs. By combining geometry-corrected initialization, progressive occlusion-aware trajectories, and confidence-guided generative optimization, P2G-GS effectively addresses multi-view misalignment and severe occlusions caused by narrow-baseline capture, outperforming state-of-the-art methods in rendering quality and robustness.

CrossView-GS: Gaussian Splatting for Cross-view Scene Reconstruction

This work presents a Gaussian Splatting method for cross-view scene reconstruction. By employing multi-branch construction with gradient-aware regularization and Gaussian supplementation, CrossView-GS effectively handles large view variations and achieves superior performance over state-of-the-art methods.

Voxel-Mesh Hybrid Representation for Real-Time View Synthesis

The Vosh framework introduces a hybrid representation that combines voxel and mesh components to achieve a flexible balance between rendering quality and speed, enabling real-time performance on mobile devices while maintaining high-quality rendering in complex regions.

DHNet: Salient Object Detection With Dynamic Scale-Aware Learning and Hard-Sample Refinement

This work introduces a dynamic scale-aware learning approach and a dense sampling strategy with graph-based feature aggregation to enhance salient object detection, effectively addressing issues in object positioning and hard-sample handling, and achieving superior performance on five benchmark datasets.

Coarse to Fine: Weak Feature Boosting Network for Salient Object Detection

This paper introduces a novel Weak Feature Boosting Network (WFBNet) for salient object detection, which enhances low-confidence regions to improve detection accuracy, particularly in complex backgrounds or with small salient objects, achieving superior performance on five benchmark datasets without post-processing.

Design and source code from Jon Barron's website