Discriminative Training of Deep Fully Connected Continuous CRFs With Task-Specific Loss

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep conditional random fields (CRFs) have advanced structured prediction in computer vision.
• Existing deep CRF models often focus on discrete labeling problems.

Purpose of the Study:

• To propose a fully connected deep continuous CRF model for both discrete and continuous labeling tasks.
• To introduce task-specific loss functions for enhanced CRF parameter learning.
• To demonstrate the model's effectiveness on semantic labeling and depth estimation.

Main Methods:

• Modeling unary and pairwise potentials using deep convolutional neural networks (CNNs).
• Jointly learning CNNs in an end-to-end fashion.
• Employing task-specific loss functions (multi-class classification loss, Tukey's biweight loss) instead of maximum likelihood estimation.
• Utilizing the closed-form solution for maximum a posteriori (MAP) inference inherent in continuous CRFs.

Main Results:

• The proposed deep continuous CRF model achieves state-of-the-art performance on multi-class semantic labeling and robust depth estimation.
• Task-specific losses enable direct optimization of MAP estimates, improving prediction quality.
• The continuous CRF model demonstrates strong performance even on discrete labeling tasks when equipped with appropriate losses.

Conclusions:

• The proposed deep continuous CRF framework offers a versatile and effective solution for structured prediction problems in computer vision.
• Task-specific loss functions are crucial for optimizing CRF parameters and achieving superior results.
• This work bridges the gap between continuous and discrete labeling by adapting a continuous CRF model for discrete tasks.