Most of the existing object detection models require large-scale datasets for training. Semi-supervised learning is becoming an important task as it reduces the work of human annotation to get a robust object detectio...
详细信息
Most of the existing object detection models require large-scale datasets for training. Semi-supervised learning is becoming an important task as it reduces the work of human annotation to get a robust objectdetection model. The generic semi-supervised learning approach is to train an initial model with a small amount of labeled data, inference the model on unlabeled data to pick up a set of new data(pseudo-labeling data) with high confidence, then train the model again with the initial labeled data and the pseudo-labeling data. Using pseudo-labeling data is one of the best methods to obtain a semi-supervised learning model. However, the initial labeled data is not always balanced and it makes the mean Average Precision (mAP) of the model lower than expected, especially comparing with other fully supervised models. In this paper, we propose a method of generating pseudo-labeling data iteratively by applying knowledge of data distribution to balance the data set and benefit the mAP. It improves the performance of a model with very few labeled examples and outperforms the generic semi-supervised learning method with only one iteration of pseudo-labeling data.
暂无评论