Few-shot Object Detection Based on Sample Balance Correction

Abstract

We propose a few-shot object detection algorithm based on sample balance correction, which introduces categorical sample balance correction and positioning sample balance correction on the basis of the Faster R-CNN network to improve the performance of few-shot detection, and categorical sample balance correction increases the passing probability of the new type pre-selection box by changing the selection rules of the pre-selection box. And the similarity calculation branch is added to the detection head to help improve the classification ability, and the positioning sample balance correction solves the sample balance in the training process and the balance problem of correcting difficult samples by improving the positioning loss function. Experiments on the PASCAL VOC dataset show that the proposed algorithm improves the detection effect when the sample size is small, up to 3.2 percentage points higher than the comparison algorithm, and has good robustness and generalization ability.