RecFRCN: Few-Shot Object Detection With Recalibrated Faster R-CNN

Abstract

Currently, Faster R-CNN serves as the fundamental detection framework in the majority of few-shot object detection algorithms. However, due to limited samples per class, the Faster R-CNN’s classification branch faces limitations in capturing specific features for each class in a few-shot scenario, leading to a bias towards false positives. These high-score false positives subsequently result in poor classification task performance. To address this issue, we propose a novel approach, named Recalibrated Faster R-CNN, which recalibrates the categories of regression boxes. Specifically, we introduce a new classification network (Rec-Net) for Faster R-CNN’s Box Predictor, including a feature extractor, a feature enhancement block (FEB), an ROI Pooling layer, and a local descriptor classifier (LDC). The feature extractor extracts features from input images, while FEB enhances these features. The ROI Pooling layer projects prediction boxes from Faster R-CNN onto a fixed-size feature map. LDC not only obtains the optimal depth local descriptors from ROI features for image-to-class measurement but also, in the few-shot setting, uses local representation as natural data enhancement for increased efficiency, ultimately enhancing the original classification scores. Our experimental results demonstrate strong performance across several benchmark tests.

Keywords

• Few-shot object detection
• metric learning
• object detection