We strive to the emerging field of 3D vision where we understand the 3D world around us. With the help of various acquisition devices, we can capture or infer a 3D model of objects, scenes, and human bodies. We can visualize the 3D model and extract semantic information to develop various applications, namely VR/AR, robotics, human augmentation, and ambient intelligence to name a few. The ultimate goal is the bridge between human and intelligent agent to benefit human.
3D information is widely used for perception, and recent trends on 3D perception utilize the state-of-the-art techniques from computer vision and machine learning. We utilize neural networks of generative models, metric learning, and/or reinforcement learning to boost the performance.
3D models are crucial for seamless AR/VR applications and realistic rendering. The key technical components include localization, pose estimation, texture and material acquisition, and lighting estimation.
The real-world 3D models are utilized for manipulation, navigation, or other robotic applications. Also, various physical interactions with 3D scenes are tightly coupled with the actions of humans or robots in addition to the physical properties of individual objects. Our focus is to build algorithms to control robot in un-constrained set-up and generate natural interaction to reflect the real-world scenarios.