Generative Landmarks

The proposed method is capable of learning landmarks for various body regions, including faces and hands. The training set consists of roughly 18k frames (about 10 minutes each) for each domain – marked and unmarked, scaled down to resolution 128 × 128.