Multi-view 3D face reconstruction with deep recurrent neural networks

作者：

Highlights：

• Recurrent neural network is effective for feature fusion in MV3DR.

• End-to-end MV3DR is achieved with a CNN and a RNN for feature extraction and fusion.

• Fusion by averaging over shapes suppresses facial identity information in 3D face.

• Fusion by aggregating facial features helps preserve facial identity information.

• Performance of 3D-aided face recognition is improved with MV3DR.

摘要

•Recurrent neural network is effective for feature fusion in MV3DR.•End-to-end MV3DR is achieved with a CNN and a RNN for feature extraction and fusion.•Fusion by averaging over shapes suppresses facial identity information in 3D face.•Fusion by aggregating facial features helps preserve facial identity information.•Performance of 3D-aided face recognition is improved with MV3DR.

论文关键词：Recurrent neural network,Long-short term memory,3D face reconstruction,Face recognition

论文评审过程：Received 15 September 2017, Revised 20 July 2018, Accepted 12 September 2018, Available online 25 September 2018, Version of Record 23 October 2018.

论文官网地址：https://doi.org/10.1016/j.imavis.2018.09.004