Universal Fine-grained Visual Categorization by Concept Guided Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing fine-grained visual categorization (FGVC) methods struggle with scene-centric and adverse viewpoint images due to reliance on object-centric assumptions.
• Mis-/over-feature activation in challenging scenarios degrades fine-grained representation learning.
• A universal FGVC framework is needed for real-world applications beyond ideal object-centric views.

Purpose of the Study:

• To develop a universal fine-grained visual categorization (FGVC) framework adaptable to real-world scenarios.
• To address limitations of current FGVC methods in scene-centric and adverse viewpoint conditions.
• To propose a novel concept-guided learning approach for robust visual categorization.

Main Methods:

• Concept Guided Learning (CGL) framework proposed, modeling categories via inherited and discriminative concepts.
• Discriminative concepts guide fine-grained representation learning through concept mining, fusion, and constraint.
• Introduced the Fine-grained Land-cover Categorization Dataset (FGLCD) with 59,994 samples to bridge dataset gaps.

Main Results:

• CGL demonstrates competitive performance on conventional FGVC tasks.
• Achieved state-of-the-art results on fine-grained aerial scenes and scene-centric street scenes.
• Showcased strong generalization capabilities in object re-identification and fine-grained aerial object detection.

Conclusions:

• Concept Guided Learning (CGL) offers a robust and universal solution for fine-grained visual categorization in diverse real-world conditions.
• The proposed CGL framework effectively overcomes challenges posed by scene-centric and adverse viewpoint imagery.
• The new FGLCD dataset facilitates further research in challenging FGVC scenarios.