Data-Efficient Matching for Object Detection with Transformer in Pin Defect Detection

Abstract

Pin is an important fixed component of high voltage line. Therefore, pin defect is an important component in the safety inspection of high voltage line. Due to the long line and harsh environment, it is difficult to obtain a large-scale dataset of pin defect. The detection with transformer (DETR) based method has achieved great success on large-size datasets, but it is less effective on small-size datasets. In this paper, we found that one of the main reasons for the poor performance of DETR-like method on small-size datasets is that the one-to-one matching of Hungarian matching leads to the small number of positive examples during training, which makes the model difficult to converge. To solve the above problems, we propose data-efficient Hungarian match (DEHM) and group object query (GOQ) to increase the number of positive examples in training. DEHM and GOQ will not add any parameters during training, and will not affect the inference speed. Extensive experiments show that DEHM and GOQ can improve the performance of DETR-like methods on both small-size dataset to achieve similar results to those on large-size datasets.