SDOF-GAN: Symmetric Dense Optical Flow Estimation With Generative Adversarial Networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Symmetric optical flow estimation is preferred for its source/target image independence.
• Current Convolutional Neural Networks (CNNs) often use asymmetric optical flow, limiting performance.
• A gap exists in applying symmetric flow estimation within state-of-the-art CNN architectures.

Purpose of the Study:

• Introduce SDOF-GAN, a novel symmetric dense optical flow model utilizing generative adversarial networks (GANs).
• Address the limitations of asymmetric flow estimation in CNN-based optical flow techniques.
• Enhance optical flow estimation accuracy and robustness through symmetry.

Main Methods:

• Developed SDOF-GAN, incorporating an inverse network for forward-backward mapping consistency.
• Employed a GAN framework where the generator estimates symmetric flow and the discriminator validates estimations.
• Utilized semi-supervised learning to leverage both labeled and unlabeled data.

Main Results:

• SDOF-GAN demonstrated significant performance improvements on five public datasets.
• The model achieved superior results compared to several representative state-of-the-art optical flow methods.
• The symmetric approach proved more effective than generic asymmetric methods.

Conclusions:

• SDOF-GAN effectively bridges the gap between symmetric flow theory and CNN implementation.
• The proposed GAN-based approach enhances the accuracy and reliability of dense optical flow estimation.
• Semi-supervised training enables efficient utilization of available data for robust model training.