Training deep retrieval models with noisy datasets: Bag exponential loss. (April 2021)

• Record Type: Journal Article
• Authors: Martínez-Cortés, Tomás; González-Díaz, Iván; Díaz-de-María, Fernando
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Image retrieval -- Noise -- Multiple instance learning -- Loss functions; 00-01 -- 99-00; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107811
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A noise-robust loss based on Multiple Instance Learning (MIL) is used to train CNNs for retrieval under noisy datasets. The Bag Exponential (BE) presented is flexible enough to be used for other purposes than dealing with noise, such as online hard positive mining. Our method allows to use noisy generated training sets, which are easy and quick to create, to adapt CNNs for image retrieval on any new object. Abstract: Although the CNNs are a very powerful tool for image retrieval, the need of training datasets properly adapted to the application at hand hinders the usefulness of such networks, specially since the datasets need to be free of noise to avoid spoiling the learning process. An ad hoc preprocessing of the dataset to mitigate the noise is a possible solution, but it is usually non-trivial and requires significant human intervention. In this paper, we pave the road for training CNNs for image retrieval with noisy datasets. In particular, we propose a novel Bag Exponential Loss function that, inspired by the Multiple Instance Learning framework, works with bags of matching images instead of single pairs, and allows a dynamical weighting of the relevance of each sample as the training progresses. The formulation of the proposed model is general enough and may serve to other purposes than dealing with noise if parameters are chosen appropriately. Extensive experimental results show the superior performance of the proposed loss with respect to the current state-of-the-art as well as its ability to cope with noisy training sets. Pytorch code available in https://github.com/tmcortes/BELoss