Weak supervision training method of pedestrian re-identification model based on micrograph learning

Abstract

The invention provides a weak supervision training method for a pedestrian re-identification model based on micrograph learning. The method comprises the following steps: firstly, grouping pedestrianpictures into bags according to shooting time periods, and distributing bag category labels; then, capturing a dependency relationship among all the pictures in each bag to generate a reliable pseudopedestrian category label for each picture in the bag of the category, and taking the reliable pseudo pedestrian category label as supervision information for pedestrian re-identification model training; then, performing integrated training of a pedestrian re-identification model and a graph model; and taking a linear combination of the graph model loss and the re-identification loss as a total loss function, and updating parameters of all layers of the network by utilizing a back propagation algorithm. According to the method, heavy manual annotation cost is not needed, and leading model performance can be achieved almost without increasing calculation complexity.

Description

technical field [0001] The invention relates to the technical field of machine vision, and more specifically, to a weakly supervised training method of a pedestrian re-identification model based on differentiable graph learning. Background technique [0002] Currently, there are three main approaches to implement the person re-identification problem: (1) extract discriminative features; (2) learn a stable metric or subspace for matching; (3) combine the above two methods. However, most implementation methods require strong supervised training labels, that is, each image in the dataset needs to be manually labeled. In addition, there are person re-identification methods based on unsupervised learning that do not require manual annotation, using local saliency matching or clustering models, but it is difficult to model significant differences across camera views, so it is difficult to achieve high accuracy. In contrast, the weakly supervised person re-identification method pr...

AI Technical Summary

Problems solved by technology

However, this patent cannot directly express graph learning as a loss function that is differentiable to network parameters, so that it can be optimized by stochastic gradient descent to achieve integrated training of graph model and pedestrian re-identification model

Images