Papers
2
Total Citations
4
H-Index
2
About
Zhi Han is a researcher at the forefront of efficient deep learning and 3D computer vision, whose work bridges the gap between theoretical model compression and practical robotic perception. His primary research areas include neural network compression, tensor decomposition, and online 3D scene reconstruction. Han’s major contributions are twofold. First, in his 2024 paper "Towards Super Compressed Neural Networks for Object Identification," he introduced a novel framework combining quantized low-rank tensor decomposition with self-attention mechanisms, achieving extreme parameter reduction for deep convolutional networks. This work directly addresses the critical challenge of deploying powerful object identification models on resource-constrained devices like mobile phones. Second, with "GeoRecon: Geometric Coherence for Online 3D Scene Reconstruction From Monocular Video," also from 2024, Han developed a method for incrementally recovering 3D meshes from monocular RGB videos. By leveraging geometric coherence, his approach enables robots to perform real-time environmental interaction tasks with significantly reduced memory consumption compared to traditional coarse-to-fine methods. Though early in his career, Han’s dual focus on model efficiency and spatial understanding positions him as an emerging leader in making advanced AI both compact and spatially aware.
Research Focus
Key Achievements
Top Papers
- 1
- 2