Instance Image Retrieval by Learning Purely From Within the Dataset

TL;DR

This paper proposes a self-supervised learning approach that uses object proposals and dataset-specific similarity information to learn effective image features for instance retrieval without relying on external datasets or labels.

Contribution

It introduces a novel dataset-specific feature learning method using self-supervision and self-boosting, bypassing the need for pre-trained models or auxiliary datasets.

Findings

Achieves competitive retrieval performance with a simple self-supervised approach.

Demonstrates effectiveness of dataset-specific learning over domain transfer methods.

Shows limitations in generalization across different datasets.

Abstract

Quality feature representation is key to instance image retrieval. To attain it, existing methods usually resort to a deep model pre-trained on benchmark datasets or even fine-tune the model with a task-dependent labelled auxiliary dataset. Although achieving promising results, this approach is restricted by two issues: 1) the domain gap between benchmark datasets and the dataset of a given retrieval task; 2) the required auxiliary dataset cannot be readily obtained. In light of this situation, this work looks into a different approach which has not been well investigated for instance image retrieval previously: {can we learn feature representation \textit{specific to} a given retrieval task in order to achieve excellent retrieval?} Our finding is encouraging. By adding an object proposal generator to generate image regions for self-supervised learning, the investigated approach can successfully learn feature representation specific to a given dataset for retrieval. This representation can be made even more effective by boosting it with image similarity information mined from the dataset. As experimentally validated, such a simple ``self-supervised learning + self-boosting'' approach can well compete with the relevant state-of-the-art retrieval methods. Ablation study is conducted to show the appealing properties of this approach and its limitation on generalisation across datasets.