On the Correlation Among Edge, Pose and Parsing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human semantic parsing, edge detection, and pose estimation are related tasks.
• Individual learning neglects contextual cue interactions.
• Prior work often fuses only two of these tasks.

Purpose of the Study:

• To develop a unified model for simultaneously learning human parsing, edge detection, and pose estimation.
• To explore effective methods for integrating pixel-level semantics, human boundaries, and joint locations.
• To investigate the benefits of joint learning over individual task training.

Main Methods:

• Proposed an end-to-end trainable Human Task Correlation Machine (HTCorrM).
• Introduced a Heterogeneous Non-Local (HNL) module to discover correlations between heterogeneous task domains.
• Utilized an asymmetric design where one task acts as the main task, supported by the other two as auxiliary tasks.

Main Results:

• HTCorrM achieved competitive performance on human parsing, pose estimation, and body edge detection.
• Designating any of the three tasks as the main task led to accuracy improvements.
• The proposed feature correlation strategy outperformed feature concatenation and post-processing methods.

Conclusions:

• Jointly learning semantic parsing, edge detection, and pose estimation in a unified model is effective.
• The HTCorrM with the HNL module successfully captures inter-task correlations.
• This approach enhances the performance of individual human-centric computer vision tasks.