I am interested in 3D computer vision and video understanding. I have a good understanding of multiple view geometry and I am passionate about Autonomous driving and AR/VR.
In particular, I have experience with:
3D computer vision:
Viewpoint estimation; 6DOF pose estimation; 3D object detection;
3D rotation estimation; Surface normal estimation; Depth estimation.
It is well-known that one-vs.-rest classification and multi-label classification need to be handled by the right classifier, i.e. softmax classifier vs. sigmoid classifiers. Could it be simpler? Yes. Quasibinary classifier might be the only classifier you need. And even better, Quasibinary classifier yields well-calibrated probabilities that are critical to tasks such as autonomous driving and clinical diagnosis.
Regression problems have always been tricky for deep neural networks to learn, mainly due to unstable gradient.
To resolve this, we proposal spherical regression, a plugin module that boosts the performance for three regression tasks in computer vision at the same time. We also present a 3D rotation estimation dataset called ModelNet10-SO3.
We present a pipeline that automatically selects and aligns textureless 3D shapes (e.g. CAD models) onto 2D images.
This is particular useful for collecting training data for autonomous driving.