Saliency for fine-grained object recognition in domains with scarce training data. (October 2019)

• Record Type: Journal Article
• Authors: Flores, Carola Figueroa; Gonzalez-Garcia, Abel; van de Weijer, Joost; Raducanu, Bogdan
• Is Part Of: Pattern recognition. Volume 94(2019:Oct.)
• Page Start: 62
• Page End: 73
• Publication Date: 2019-10
• Subjects: Object recognition -- Fine-grained classification -- Saliency detection -- Scarce training data; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.05.002
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We investigated the role of saliency on improving the classification accuracy when the training data is scarce. We considered adding a saliency branch to an existing CNN architecture (AlexNet, ResNet-50 and ResNet-152). We validated our approach on the fine-grained object recognition problem. Experimental results confirmed that our approach is useful for the case when the available training data is scarce. Our experiments show that there exists a clear correlation (Pearson coefficient) between the performance of saliency methods on standard saliency benchmarks and the performance gain that is obtained when incorporating them in a object recognition pipeline. Abstract: This paper investigates the role of saliency to improve the classification accuracy of a Convolutional Neural Network (CNN) for the case when scarce training data is available. Our approach consists in adding a saliency branch to an existing CNN architecture which is used to modulate the standard bottom-up visual features from the original image input, acting as an attentional mechanism that guides the feature extraction process. The main aim of the proposed approach is to enable the effective training of a fine-grained recognition model with limited training samples and to improve the performance on the task, thereby alleviating the need to annotate a large dataset. The vast majority of saliency methods are evaluated on their ability to generate saliency maps, and not on their functionality in a complete vision pipeline. Our proposed pipeline allows to evaluate saliency methods for the high-level task of object recognition. We perform extensive experiments on various fine-grained datasets (Flowers, Birds, Cars, and Dogs) under different conditions and show that saliency can considerably improve the network's performance, especially for the case of scarce training data. Furthermore, our experiments show that saliency methods that obtain improved saliency maps (as measured by traditional saliency benchmarks) also translate to saliency methods that yield improved performance gains when applied in an object recognition pipeline.