Unifying the Visual Perception of Humans and Machines on Fine-Grained Texture Similarity
Weibo Wang (Ocean University of China), Xinghui Dong (Ocean University of China)*
The 33rd British Machine Vision Conference
Abstract
Texture similarity plays an important role in texture analysis and material recognition. However, the prediction of perceptually-consistent fine-grained texture similarity is a challenging task. It has been found that the discrepancy between the texture representation methods and the similarity metrics utilised by humans and algorithms should account for the dilemma. To address this problem, we propose a novel Perceptually Motivated Texture Similarity Prediction Network (PMTSPN), which comprises a siamese Conformer with multi-scale bilinear pooling (SC-MSBP) and a metric learning network (MLN). The SC-MSBP learns a texture representation capturing the Higher Order Statics (HOS) in different spatial scales, while the MLN learns a similarity metric from the features which encodes the short-range, long-range and lateral interactions. The PMTSPN can be trained using a set of human perceptual similarity data. Our results show that the PMTSPN produces the more consistent similarity prediction with human perception, compared with its counterparts. We attribute the promising performance to both the powerful texture representation and the effective similarity metric learnt by the PMTSPN.Video
Citation
@inproceedings{Wang_2022_BMVC,
author = {Weibo Wang and Xinghui Dong},
title = {Unifying the Visual Perception of Humans and Machines on Fine-Grained Texture Similarity},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0839.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by