Multi-Instance Dynamic Ordinal Random Fields for Weakly-supervised Facial Behavior Analysis.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Weakly-supervised learning problems often involve training data with labels only at the bag level.
• The Multi-Instance Dynamic-Ordinal-Regression (MI-DOR) setting specifically deals with ordinal instance labels within temporal sequences.
• Existing methods may not fully capture the complex dependencies in such data.

Purpose of the Study:

• To propose a novel Multi-Instance-Learning (MIL) framework for the MI-DOR setting.
• To develop a method that models temporally-dependent latent instance labels within an Undirected Graphical Model.
• To extend the framework for partially-observed MI-DOR problems, reducing annotation effort.

Main Methods:

• Introduced Multi-Instance Dynamic Ordinal Random Fields (MI-DORF) using an Undirected Graphical Model.
• Incorporated high-order potentials to model various MIL assumptions within the energy function.
• Extended the framework to handle partially-observed instance labels during training.

Main Results:

• MI-DORF significantly outperforms alternative learning approaches on weakly-supervised facial behavior analysis tasks.
• Demonstrated effectiveness in Facial Action Unit (DISFA dataset) and Pain (UNBC dataset) intensity estimation.
• Showcased the framework's ability to substantially reduce data annotation efforts.

Conclusions:

• The proposed MI-DORF framework offers a powerful solution for weakly-supervised learning in MI-DOR settings.
• MI-DORF provides state-of-the-art performance in complex tasks like facial behavior and pain intensity analysis.
• The method effectively reduces the need for extensive data labeling, making it highly practical.