How visual chirality affects the performance of image hashing

Abstract

Visual chirality reveals the phenomenon that chiral data will present different semantics after flipping. Although image flipping is widely used in image hashing learning as a data augmentation technique, the effect of learning chiral image data on hashing performance has not been fully discussed. To explore this issue, this paper first designs an approach to recognize images with chiral cues, then constructs the chiral datasets including different proportions of images with chiral cues, and finally analyzes and discusses the performance change via testing three representative image hashing methods with different hash code lengths on constructed chiral datasets. In addition, to understand the effect of visual chirality from an internal perspective, we illustrate visual results of activated regions between some original images with chiral cues and their flipped ones. We conduct the above experiments on three public image datasets including VOC2007, MS-COCO, and NUS-WIDE. Experimental results reveal that different proportions of chiral data will greatly affect the performance of image hashing and the best performance appears when the proportion of images with chiral cues accounts for 15% $\sim$25% or 75% $\sim$85%. The code of this work is released at: https://github.com/lzHZWZ/Visual_Chirality_Hashing.