Semi-supervised Learning Method for Object Detection based on Adjacent Frame Consistency Measurement

Deep neural network requires a large amount of labeled data to train the parameters of the model, but the cost of data labeling is high, and only a small part of the collected data is labeled as the dataset. Therefore, a large amount of unlabeled data is not used in the training process. The reasonable utilization of unlabeled data has attracted much attention from the academic and industrial areas. In this paper, a semi-supervised learning method for object detection based on the consistency measurement of adjacent frames was proposed. The semi-supervised learning task of the object detection model was completed by using the semantic similarity of the adjacent frames in the video. We found that the adjacent frames of the video have similar semantic properties and meet the requirements of consistency samples. Using adjacent frames as consistency samples to generate the loss on unlabeled data, the average precision of the neural network could get 83.0%.