Multi-source domain data set joint training REID model method, device and medium

Abstract

The invention discloses a method and device for jointly training an REID model through multi-source-domain data sets and a medium. The method comprises the steps of training a to-be-trained multi-layer perceptron in a to-be-trained depth dynamic convolution kernel based on a plurality of REID source domain data sets, so that the converged multi-layer perceptron can determine domain information and camera information corresponding to each REID source domain data set based on sample information of the REID source domain data sets; based on the sample information, the domain information and the camera information corresponding to each REID source domain data set, training a to-be-trained dynamic convolution kernel in the to-be-trained depth dynamic convolution kernel to determine a dynamic convolution kernel corresponding to each REID source domain data set; and generating a depth dynamic convolution kernel based on the multilayer perceptron and the dynamic convolution kernel, and replacing the standard convolution in the to-be-trained REID model with the depth dynamic convolution kernel to generate the REID model. According to the method, the pedestrian re-identification model is trained in combination with the multi-source domain data set.

Description

technical field [0001] The present application relates to the technical field of person re-identification, and in particular, to a method, device and medium for jointly training a REID model with multi-source domain data sets. Background technique [0002] At present, in the field of person re-identification, the effect of the model trained with the dataset of a single source domain is better than that of the joint dataset of multiple source domains, which is a kind of domain conflict problem. Person Re-ID (ReID) datasets from different source domains come from different regions and times, resulting in great differences in appearance among the samples of Person Re-ID (ReID) datasets from different source domains. Due to this domain conflict, traditional methods cannot process data from different source domains at the same time, especially when the data distribution of these source domains is very different, so the trained model cannot recognize target images from non-trainin...

AI Technical Summary

Problems solved by technology

Due to this domain conflict, traditional methods cannot process data from different source domains at the same time, especially when the data distribution of these source domains is very different, so the trained model cannot recognize the target image of the non-training source domain.

Images