Method and system for object re-identification

Abstract

The disclosure relates to a method and a system for object re-identification, the method comprises: extracting, by a convolutional neural network, features for object images in a plurality of datasets, wherein the convolutional neural network comprises a plurality of neurons, and each of the features is composed of neuron responses corresponding to each of the neurons; determining, for each of thedatasets, a neuron impact score for each of the neurons according to the extracted features; updating, for each of the datasets and according to the determined score, the features by adjusting a weight for each of the neuron responses in the features; training the convolutional neural network based on the updated features; extracting, by the trained convolutional neural network, a plurality of first features for a target object image and a plurality of second features for each of object images in a given image set, respectively; and locating the target object in the given image set accordingto the extracted first and second features.

Description

technical field

[0001] The present disclosure relates to a method and system for object re-identification. Background technique

[0002] In computer vision, a domain usually refers to a dataset whose samples follow the same distribution as the underlying data. Multiple datasets with different data distributions are often proposed to solve the same or similar problems. Multi-domain learning aims to solve problems related to datasets across different domains simultaneously by using all the data from these datasets. The advent of large-scale training data has contributed to the success of deep learning, however, it also introduces interesting problems in multi-domain learning. Many studies have shown that fine-tuning deep models pre-trained on large-scale datasets is effective for other related domains and tasks. However, in many specific domains, such large-scale datasets for learning robust and general feature representations do not exist.

[0003] Another interesting asp...

Images