Xiao-Ming Wu 伍晓鸣

M.S. student

School of Computer Science and Engineering

Sun Yat-sen University

Email: wuxm65@mail2.sysu.edu.cn

[G o o g l e   Scholar]    [GitHub]

Biography

I'm currently a third-year master student at Sun Yat-sen University, advised by Prof. Wei-Shi Zheng, where I develop the scientific ability and a good taste of it. Now I am also a visiting student in Macquarie university, enjoying an interesting research journey with Prof. Longbing Cao. Previously, I obtain my B.E. degree in Shandong University. At that time, I had my first attempt on scientific research and cultivate the interest in it, advised by Prof. Xin-Shun Xu and Prof. Xin Luo.

Research is for curiosity and fun.

I am interested in general deep learning topics, including algorithms, theories, systems and its applications on vision, language scenarios. My long-term goal is to build a strong, efficient (reach human-level), and safe (surpass human level) AI agent. Now I am mainly focusing on embodied AI. Previously, I also conducted research on model compression and hashing retrieval.

News

Selected Publications

Below are my selected publications. (& means equal contribution, * refers to corresponding author.)

(Dexterous Grasp) We propose a novel task "Dexterous Grasp as You Say" (DexGYS), with a benchmark and a framework, enabling robots to perform dexterous grasping based on human commands expressed in natural language.

(6-DoF Grasp) We propose a Real-to-Sim framework for 6-DoF Grasp detection, named R2SGrasp, with the key insight of bridging this gap in a real-to-sim way, and build a large-scale simulated dataset to pretrain our model to achieve great real-world performance.

(Privacy Protection) We propose a two-step iterative method (PixelFade) for person privacy-preserving, with the partial replacement step to turn image into noise to resist recovery attack, and the constrain operation step to maintain Re-id semantics.

(6-DoF Grasp) We propose a new economic grasping framework for 6-DoF grasp detection to economize the training resource cost and meanwhile maintain effective grasp performance.

(Image Restoration) We propose a diffusion-based universal image restoration model, with an assemble-then-separate (like the hourglass) mapping for multi-task training, to learn the shared information between different tasks.

(Binary Neural Networks) We propose a new perspective to view the binary neural network training: equilibrium between estimating error and gradient stability, and design a simple and effective gradient estimator (ReSTE) to balance it well.

(Hashing Retrieval) We propose a novel cross-modal online hashing method, directly exploiting the similarity between newly coming data and old existing data in the Hamming space to learn discriminate hashing codes.

(Hashing Retrieval) We explore a new task, incremental multi-modal hashing, and introduce semantics to solve this task, handling the dimension mismatching problem and mitigating the inconsistent problem that occurs when new classes come.

Services and Activities

Selected Awards