Unsupervised cross-domain pedestrian re-identification method based on clustering

Abstract

The invention discloses an unsupervised cross-domain pedestrian re-identification method based on clustering, and belongs to the field of computer vision. The method specifically comprises the following steps: (1) establishing a neural network for pre-training; (2) inputting the labeled data in the source domain into a pre-training network, and performing supervised pre-training; (3) building a double-flow mutual learning network framework for fine adjustment; (4) training a double-flow mutual learning framework, and outputting a pedestrian re-identification model suitable for a target domain.Unsupervised cross-domain pedestrian re-identification is realized by utilizing a labeled source domain data set, a label-free target domain data set and a mutual learning network framework; according to the method, a self-attention mechanism and global joint pooling operation are introduced, a new loss function, namely joint flexible optimization loss, is provided, and a clustering method more suitable for open set data is selected, so that the model performance is obviously improved.

Description

Technical field [0001] The invention belongs to the technical field of image retrieval, and specifically designs an unsupervised cross-domain pedestrian re-recognition method based on clustering for cross-domain image retrieval. Background technique [0002] Pedestrian re-identification (re-ID) is a sub-problem of image retrieval. Its purpose is to determine whether the pedestrians in the images taken by different cameras are the same. This is a promising but challenging direction, because human images are usually affected by factors such as occlusion, illumination, and posture changes, making re-identification difficult. With the advent of convolutional neural networks and residual neural networks, pedestrian re-recognition based on supervised learning has achieved high performance. However, supervised learning requires a large amount of labeled pedestrian data, and due to the existence of domain gaps, the model trained on the source data set often suffers a great performance d...

AI Technical Summary

Problems solved by technology

[0005] (2) The number of classes in the target domain is uncertain, and the clustering algorithm will produce some wrong labels, and noise will inevitably appear in the final pseudo-labels;

[0006] (3) Using the k-means clustering method needs to manually set the number of clusters, and different values ​​of the number of clusters have a great impact on the clustering effect, and unsupervised cross-domain person re-identification is an open set task, we The number of categories in the target domain is unknown

Images