Adaptive Feature Learning for Unbiased Scene Graph Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene Graph Generation (SGG) aims to identify objects and their relationships.
• Existing methods struggle with feature training indeterminacy and label confusion.
• Predicate representation learning is crucial for SGG accuracy.

Purpose of the Study:

• To enhance predicate representation learning in Scene Graph Generation.
• To address challenges from indeterminate feature training and label confusion.
• To improve the accuracy and robustness of SGG models.

Main Methods:

• Proposed an Adaptive Message Passing (AMP) network for dynamic information propagation and de-noising.
• Introduced a feature-assisted training paradigm to guide predicate feature learning.
• Developed a Bi-level Curriculum learning scheme (BiC) to combat long-tailed distributions and label interference.

Main Results:

• The proposed AMP-BiC method demonstrated superior comprehensive performance on multiple SGG datasets.
• AMP effectively generated discriminating representations for neighbor nodes.
• BiC successfully preserved distinct predicate representations while resisting biased predictions.

Conclusions:

• The AMP-BiC method significantly enhances Scene Graph Generation by improving predicate representation learning.
• The novel approach effectively tackles label confusion and data bias issues.
• Results confirm the method's effectiveness and superior performance in SGG tasks.