Core Concepts
Proposing a point-level weakly-supervised expression spotting framework with multi-refined pseudo label generation and distribution-guided feature contrastive learning to enhance performance.
Abstract
The article introduces a weakly-supervised expression spotting framework to address the limitations of frame-level and video-level labeling methods. It proposes innovative strategies, MPLG and DFCL, to improve pseudo label generation and feature contrastive learning. Extensive experiments on three datasets demonstrate promising results comparable to fully-supervised methods.
Introduction to facial expressions as non-verbal communication.
Two primary stages of expression processing: spotting and recognition.
Existing spotting models categorized into frame-level and video-level labeling-based methods.
Proposal of a point-level weakly-supervised expression spotting framework (PWES).
Strategies MPLG and DFCL designed to enhance model training and feature similarity.
Experimental results on CAS(ME)2, CAS(ME)3, SAMM-LV datasets showing promising performance.
Stats
Extensive experiments on CAS(ME)2, CAS(ME)3, SAMM-LV datasets demonstrate PWES achieves promising performance comparable to that of recent fully-supervised methods.
Quotes
"MEs are subtle, unconscious facial movements that typically last less than 0.5 second."
"Analyzing MEs is particularly valuable in high-stakes situations such as business negotiation."