Proposal Distribution Calibration for Few-Shot Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot object detection (FSOD) faces challenges adapting models to novel classes with limited data.
• The standard two-step training paradigm involves pre-training on base classes and fine-tuning on novel classes.
• Biased region proposal networks (RPNs) in the pre-training phase hinder performance on novel classes due to suppressed unlabeled instances.

Purpose of the Study:

• To address the proposal distribution bias in FSOD.
• To enhance the localization and classification capabilities of the region of interest (RoI) head for novel classes.
• To improve the adaptation of object detectors to low-data scenarios.

Main Methods:

• Introduced a proposal distribution calibration (PDC) approach.
• Sampled proposals based on base proposal statistics to correct distribution bias.
• Applied additional localization and classification losses to sampled proposals for semantic fine-tuning.

Main Results:

• PDC effectively enhances the RoI head's abilities by leveraging base training localization.
• The method enriches high-quality positive samples, facilitating semantic fine-tuning.
• Achieved state-of-the-art performance on Pascal VOC and MS COCO datasets for FSOD.

Conclusions:

• PDC is a simple yet effective method for improving few-shot object detection.
• The approach successfully mitigates proposal distribution bias, leading to better performance on novel classes.
• The proposed technique offers a promising direction for adapting object detectors in low-data regimes.