Qian Zheng 郑 乾

Assistant Professor

College of Computer Science and Technology,
Zhejiang University, Hangzhou, China.
Email: qianzheng@zju.edu.cn
Official Academic Profile: [En][中文][CV (Jul 2025)]

• Mitigating the impact of noise and sparsity brought by event streams for event-guided novel view-time synthesis

• Exploit spiking neural networks' temporal heterogeneity through logit gradient analysis
• Highest reported SNN accuracy (up to Jun 2025): 85.83% (ImageNet-1K) and 87.63±0.15% (CIFAR10-DVS)

• Event-enhanced dynamic 3DGS model taking the input of motion-blurred monocular video
• Reconstruct the motion mask field and separate static and dynamic regions

• Event-Guided space-time video super-resolution with fewer adjacent frames
• Excel in handling large motion scenarios

• Ensure consistent brightness for HDR video reconstruction
• Achieve LDR-to-HDR transformation with single-exposure LDR frames

• Reconstruct HDR radiance field even if input images are degraded and are single-exposured
• Reconstruct Camera Response Function (CRF) from single-exposure images

• A unified framework for SNN robustness analysis
• An inherentrobust SNN method interoperable with other methods, such as adversarial training

• Spiking neuron to reduce excessive Gaussians
• Faster, lower storage, and higher accuracy 3DGS-based surface reconstruction
• Stable performance superiority of bio-inspired neural computation over artificial neuron-based computation on core AI tasks

• Event-based model establishes relationship between events and intrinsic components
• Multi-view consistency of events to extract specular-related clues
• Event-guided intrinsic decomposition framework enables relighting under extreme conditions

• An ANN-SNN hybrid MLP for NeRF-based 3D surface reconstruction
• Demonstrate the superiority of spiking neuron on 3D surface representation
• Stable performance superiority of bio-inspired neural computation over artificial neuron-based computation on core AI tasks

• An ANN2SNN conversion method with high performance and low time step
• Highest reported SNN accuracy (up to Jun 2025): object detection (mAP: 73.43%, time steps: 7, PASCAL VOC 2007), and semantic segmentation (mIoU: 69.7%, time steps: 15, PASCAL VOC 2012)

• Alleviate the semantic gap within known and unknown semantic subspaces
• A generalized fMRI-to-image reconstruction method that adaptively weights the semantic and structural information

• Address point cloud registration problem under scenarios of low overlap and photometric inconsistency
• An unsupervised solution based on 3DGS for point cloud registration

• Answer three fundamental questions: 1) What is the desired deep feature? 2) How to resolve PS challenges? 3) What is the desired network architecture?
• A solution based on answers that achieves SOTA performance on several PS benchmarks

• A novel setup for Natural Light Uncalibrated Photometric Stereo
• A light prior that leverages object's occluding boundaries for reliable environment light

• An interesting idea of the proposed irradiance fields
• Irradiance fields can be integrated into and jointly optimized with radiance fields to recover the geometry and HDR of a panoramic image

• A differentiable shadow handling method and an anisotropic reflectance model
• A self-supervised framework that simultaneously optimizes shape, anisotropic reflectance, shadow map, and light conditions

• Identify 3 problems in the PF-ODEs-based score distillation method by connecting consistency distillation to score distillation
• An improved score distillation method to enhance details and fidelity of generated 3D assets

• An arbitrary Style Image Manipulation Network to achieve zero-shot style image manipulation
• Balance computational efficiency and high resolution

• Enable the detection of whether a response is OOD for the reward model.
• The first method that uses value regularization to address reward over-optimization and penalizes OOD values without affecting ID ones

• A method to estimate finite-horizon non-discounted constraints in Safe RL tasks
• A deep Safe RL algorithm to address tasks with finite-horizon constraints

• Construct multiplier-penalty function that dampens cost oscillation for stable convergence while being equivalent to the primal constrained problem to precisely control safety costs
• Provide an implementation based above construction