GCSTG: Generating Class-confusion-aware Samples with a Tree-structure Graph for Few-shot Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-Shot Object Detection (FSOD) struggles with novel classes due to limited data.
• Poor class hierarchy construction leads to inter-class confusion, degrading FSOD performance.
• Intra-super-class confusion, misclassification within broader categories, is a key challenge.

Purpose of the Study:

• To develop a method for constructing precise class hierarchies in FSOD.
• To mitigate both inter-class and intra-super-class confusion in novel object detection.
• To enhance the performance of existing FSOD methods.

Main Methods:

• Generation of class-confusion-aware samples using a tree-structure graph.
• Injecting noise into samples to maximize confusion category confidence.
• Implementing a confusion-aware curriculum learning strategy for gradual training.

Main Results:

• The proposed method effectively addresses intra-super-class confusion.
• Generated samples aid in building more accurate class hierarchies.
• The approach serves as a plug-in, consistently improving FSOD model performance.

Conclusions:

• The novel sample generation and curriculum learning strategy significantly enhance FSOD.
• This method offers a robust solution to class confusion problems in few-shot learning.
• The approach demonstrates broad applicability and performance gains across various FSOD models.