Cost Function Unrolling in Unsupervised Optical Flow.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Steepest descent algorithms in deep learning rely on gradients, which are difficult to compute for complex or non-differentiable cost functions.
• Current solutions often involve increasing deep neural network (DNN) model size and complexity, leading to higher computational costs.
• Numerical instability and challenges in gradient computation near singular points hinder the training of DNNs.

Purpose of the Study:

• To introduce a novel mechanism, Cost Unrolling, to enhance DNNs' ability to solve complex cost functions without architectural changes or increased computational load.
• To develop an iterative differentiable alternative to the Total Variation (TV) smoothness constraint.
• To improve gradient stability, convergence speed, and prediction accuracy in DNNs.

Main Methods:

• Proposed a mechanism named Cost Unrolling to improve DNN performance on complex cost functions.
• Derived an iterative differentiable alternative to the Total Variation (TV) smoothness constraint.
• Integrated the novel loss function into DNN training for tasks such as image denoising and unsupervised optical flow.

Main Results:

• The proposed differentiable TV alternative resulted in more stable gradients during DNN training.
• The method enabled faster convergence compared to traditional approaches.
• Significant improvements were observed in image denoising and unsupervised optical flow tasks, particularly in predicting flows at occluded regions and achieving sharper motion boundaries.

Conclusions:

• Cost Unrolling offers an effective way to enhance DNNs without increasing model complexity or computational requirements.
• The differentiable TV constraint alternative leads to improved training dynamics and superior performance in vision tasks.
• The method shows particular promise for improving optical flow prediction in challenging scenarios like occluded regions.