An efficient encoder-decoder model for portrait depth estimation from single images trained on pixel-accurate...

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Depth estimation from single images is vital for applications like AR and autonomous driving.
• Acquiring dense ground-truth depth data is challenging, limiting training dataset diversity.
• Existing methods struggle with accuracy due to data limitations.

Purpose of the Study:

• To develop a robust synthetic data generation pipeline for facial depth estimation.
• To evaluate state-of-the-art depth estimation algorithms using synthetic data.
• To propose an improved neural network for enhanced facial depth estimation.

Main Methods:

• A 3D pipeline generating 100 synthetic virtual human models for 2D facial images and depth data.
• Evaluation of existing single-image depth estimation algorithms on the synthetic dataset.
• Development of an efficient encoder-decoder neural network trained with synthetic data and combined loss functions.

Main Results:

• The synthetic dataset provides controllable variations for training depth estimation models.
• The proposed neural network outperforms current state-of-the-art methods on four public datasets.
• The synthetic data approach yields more reliable ground truth for depth estimation.

Conclusions:

• The novel synthetic data pipeline and improved neural network establish a new state-of-the-art in facial depth estimation.
• This approach overcomes limitations of real-world data acquisition for training depth estimation models.
• The method demonstrates superior performance and computational efficiency for single-image facial depth estimation.