Towards Generalized Few-Shot Open-Set Object Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open-set object detection (OSOD) is crucial for real-world applications but struggles with limited data.
• Existing OSOD methods often fail in few-shot scenarios, leading to misclassification of unknown objects as known ones.

Purpose of the Study:

• To address the challenge of generalized few-shot open-set object detection (G-FOOD) in data-scarce environments.
• To develop a method that maintains few-shot detection performance while improving the rejection of unknown objects.

Main Methods:

• Proposed the Few-shOt Open-set Detector (FOOD) algorithm for G-FOOD.
• Introduced a novel class weight sparsification classifier (CWSC) to prevent overfitting on known classes.
• Implemented a novel unknown decoupling learner (UDL) to create a distinct decision boundary for unknown objects.

Main Results:

• The CWSC reduces co-adaptability between classes, mitigating overfitting.
• The UDL enables confident identification of unknown objects without thresholds or prototypes.
• The FOOD method improved unknown class F-scores by 4.80%-9.08% across various shots on the VOC-COCO dataset compared to state-of-the-art OSOD methods.

Conclusions:

• The proposed FOOD algorithm effectively tackles the G-FOOD problem in few-shot settings.
• The method demonstrates significant improvements in detecting known objects and rejecting unknown ones, outperforming existing approaches.