Fixing Background Misclassification in Few-Shot Object Detection via Product of Experts.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) struggles with limited data, leading to poor object representation.
• Two-stage fine-tuning transfers knowledge but often misclassifies novel objects as background due to domain gaps.

Purpose of the Study:

• To develop a robust framework for few-shot object detection that mitigates background misclassification.
• To improve the accurate identification of novel object categories with limited labeled data.

Main Methods:

• Proposes a Product of Experts (PoE) formulation to estimate joint distributions over background and novel categories.
• Combines unnormalized logits from independently trained classifiers without modifying base models.
• Introduces a strategy for identifying additional novel-category instances within the base dataset to augment fine-tuning data.

Main Results:

• The proposed method significantly reduces misclassification of novel objects as background.
• Achieves consistent improvements across various baselines in few-shot object detection.
• Demonstrates state-of-the-art performance on PASCAL VOC and COCO datasets.

Conclusions:

• The PoE-based framework effectively addresses a key limitation in two-stage FSOD pipelines.
• The approach is architecture-agnostic, computationally efficient, and integrates seamlessly with existing methods.
• Offers a significant advancement for accurate object detection with scarce labeled data.