February 10, 2020
Guidelines:
Pose-invariant, subject-independent, expression-discriminative representation learning
Experiments:
-
baseline: directly regress 3DMM expression coefficients
-
multi-task learning: jointly regress 3DMM expression coefficients and 2D facial landmarks
Optimization:
3DMM coefficients adaptively weighted, triplet loss/ contrastive loss, face recognition
Choose the appropriate architecture (ResNet 50 / face analysis network / which layers to share) and hyper-parameters to demonstrate the effectiveness of Multi-Task Learning, i.e. each component contributes to the overall performance.
Goal:
ACM Multimedia Conference 2020
-
Regular Papers Submission (Abstract) 21 Mar 2020
-
Regular Papers Submission 28 Mar 2020