Person Retrieval And Model Compression Based On Deep Learning

Given the feature maps of a pedestrian image,there often exist strong correlations between the features within the pedestrian body part and weak correlations between the features from the pedestrian region and the background.Although multi-granularity-based models perform well on existing popular datasets,they lack the ability of handling the background,part missing,changing viewing perspectives,etc.,since the relation between granularities is ignored.In this thesis,a regional-correlation-based method is proposed for pedestrian re-identification.The proposed method integrates the information of regional correlations into the pedestrian feature representation by a context information processing module,which enhances the identification performance.Two improvements for person re-identification are proposed in this thesis.Firstly,the CNN model is trained with stochastic depth,where the model randomly skips some convolutional layers.This can greatly enhance the generalization ability of each layer.Secondly,in order to handle the problem of occlusion on pedestrian bodies,a novel dropblock is proposed,in which the occlusion rate of the featuremap will decay linearly with depth.Finally,the model is also compressed with the student-teacher knowledge distillation method,transferring the knowledge from the original model to a small-sized model.To validate the proposed method,the model is tested with both the single-domain and cross-domain manners on four popular datasets.Rank-1 accuracy and mean average precision(mAP)are used as the evaluation metrics.Densely conducted experiments have demonstrated the effectiveness of the context information processing module.By visualizing the features,it is easily to notice the ability of soft detection and feature invariance of the context information processing module.Moreover,the enhancement of the generalization ability introduced by the two proposed improvements is also demonstrated by our ResNet-based experiments on theMarket1501 dataset.Lastly,knowledge distillation is also supported by the experimental results,by which a better-performed student model can be obtained.