Few-Shot Object Detection With Self-Supervising and Cooperative Classifier.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) aims to identify novel object categories with limited training data.
• Existing FSOD methods often overlook crucial image structural and semantic information.
• A key challenge is misclassification between data-rich base classes and data-scarce novel classes.

Purpose of the Study:

• To propose a novel FSOD approach addressing performance degradation in novel classes.
• To mitigate misclassification errors, particularly novel classes being mistaken for base classes.
• To enhance the learning of structural and semantic image features for improved detection.

Main Methods:

• Introduced a Few-Shot Object Detection with Self-Supervising and Cooperative Classifier (FSOD-SSCC) approach.
• Employed double RoI heads within Fast-RCNN to learn specialized features for novel classes.
• Integrated self-supervised learning (SSL) for richer structural and semantic feature extraction.
• Developed a cooperative classifier (CC) with base-novel regularization to maximize class separability.

Main Results:

• Identified false-positive samples, specifically novel classes misclassified as base classes, as a primary performance bottleneck.
• The proposed FSOD-SSCC method demonstrated superior performance compared to state-of-the-art baselines.
• Achieved significant improvements on benchmark datasets like PASCAL VOC and COCO.

Conclusions:

• The FSOD-SSCC approach effectively enhances few-shot object detection capabilities.
• Leveraging self-supervised learning and a cooperative classifier significantly reduces misclassification errors.
• The method shows strong generalization and outperforms existing techniques on standard datasets.