I am enthusiastic about data-efficient learning by pre-wiring deep networks with generic innate priors. My research is about re-parameterizing old-school feature engineering for end2end learning, where the inductive knowledge no longer needs to be learned from big data. I have been working on a variety of tasks on 2D/3D scene parsing, e.g., detecting line segments, vanishing points, and symmetry planes from single-view images. The ultimate goal is to better understand the 3D world without relying on large manually labeled datasets. I also have experience in vision for industrial inspection.
Please view my personal homepage here.