Image-to-Image Translation For Semi-supervised Pedestrian Detection

The purpose of pedestrian detection is to distinguish the pedestrians from the background correctly given an image or a clip of video.There are general and practical meanings while a great number of related applications have been developed such as surveillance system or automatic pilot.Recently,neural networks are getting more and more popular among the researchers.The performance of neural networks has outperformed those handcraft-feature based methods.Nevertheless,the successful training of a neural network requires a lot of labeled data which is expensive in reality.This work attempts to use only a little labeled data and exploit a great amount of unlabeled data to promote the performance of the model under the setting of semi-supervised learningIn order to alleviate the insufficiency of labeled data,the high-confidence samples are collected as pseudo-labeled data produced by a pre-trained detector trained on a little labeled data.Unfortunately,such pseudo-labeled data is considered unreliable since the performance of the pre-trained detector is unsatisfactory.In addition,the unreliable pseudo-labeled data will mislead the training of the detector.In this work,we propose an image-to-image translation model in a fashion of adversarial learning to translate an unreliable instance to the reliable one for data augmentation.Two regularization terms,namely semantic similarity and appearance consistency,are proposed to guarantee the quality of the translated instances Then a classifier is trained over the labeled data and the translated instances to produce more reliable pseudo-labeled data.Finally,the detector can be re-trained using those pseudo-labeled data.We evaluate the effectiveness of our model on three benchmark datasets,advancing state-of-the-art performance for semi-supervised pedestrian detection.