RecFRCN: Few-Shot Object Detection With Recalibrated Faster R-CNN
Youyou Zhang, Tongwei Lu
Abstract
Currently, Faster R-CNN is used as the fundamental detection framework in the majority of few-shot object detection algorithms. However, the under-explored few-shot classification branch generates a huge number of low-quality scores, and this high-score false positive leads to poor classification task performance. In order to solve this problem, we propose a novel method, called Recalibrated Faster R-CNN, to recalibrate the categories of regression boxes. Specifically, we introduce a new classification network (Rec-Net) for Faster R-CNN’s Box Predictor, consisting of a feature extractor, a feature enhancement Block (FEB), an ROI Pooling layer, and the local descriptor classifier (LDC). The feature extractor extracts features from input images. FEB enhanced the features obtained by the feature extractor. The ROI Pooling layer projects each prediction box output from Faster R-CNN into the feature map and pools it into a feature map with a fixed size. LDC not only obtains the optimal depth local descriptor from each ROI features for the image-to-class measure, but also, in the case of the few-shot setting, since the adoption of local representation can be seen as a natural data enhancement, this measure can be more efficient and thus improve the original classification score. The experimental results show that our method gets good scores on several benchmark tests.