תקציר יחידת-הוראה
-
Title: What everyone needs to know about 3D facesProf. Tal HassnerSunday 31/5/20Abstract:What is the best quality 3D face shape reconstruction one can get from a single, low quality face photo? In particular, how well can you reconstruct a face if it is (partially) occluded? And how do you even measure the quality of a 3D face reconstruction? In this talk I will answer these questions and others. I will share some surprisingly simple yet highly effective methods for computing 3D face shape and position from single images. In particular, I will explain why data collection and annotation for these tasks can be easy and almost effortless. I will expose widely held misconceptions on the use of facial landmark detection methods, the role of 3D face shapes, and the metrics used to evaluate reconstructed face shapes and face pose estimates. Finally, as a recurring theme in this talk, I will claim that old-school, pre-deep learning methods can still play an important role, even for developing modern, state of the art, deep learning solutions.