Semi-supervised Human Detection Via Region Proposal Networks Aided By Verification

Pedestrian detection is an important issue in the field of computer vision.Related technologies have been widely used in many fields,such as driverless,intelligent video surveillance and so on.Training effective pedestrian detection models usually requires a large number of labeled samples.However,there are usually differences from real-life scenarios data to standard data.It is unrealistic to manually label plenty of training samples for each application scenario.When the detection models trained by standard datasets are applied to real-life scenarios data,the performance usually drops.This article focuses on the problem of semi-supervised pedestrian detection,making full use of limited labeled samples and a large number of unlabeled samples to effectively enhance pedestrian detector performance in the real-life scene.This paper first introduces the consistency regularization term to constrain the feature expression of two adjacent candidate windows to improve the robustness of the network.Secondly,the proposed modified region proposal network is trained on a limited manually annotation data to obtain an initial model for pedestrian detection.In order to expand the annotation data,the initial detection model is applied to the unlabeled data,and the detection result with high confidence is collected as a pseudo-labeling samples.Thirdly,for the different error types in high confidence candidate samples,a verification model is designed to effectively evaluate the distribution of foreground regions in on candidate samples.Finally,we re-train the proposed modified region proposal network.We propose that the consistency regularization term can improve the generalization capacity of the region proposal network.The verification model based on the saliency analysis can improve the quality of pseudo-labeling sample and control of error propagation.The training framework based on self-paced learning can gradually expand the training dataset,and finally improve the detector's performance in the real-life scene.This paper also validates the effectiveness of our main components on various pedestrian detection standard dataset,and compares with the state-of-the-art semi-supervised pedestrian detection methods.