BFC-BL: Few-Shot Classification and Segmentation combining Bi-directional Feature Correlation and Boundary constraint
Haibiao Yang (Guangdong University of Technology),* Zeng Bi (Guangdong University of Technology), Pengfei Wei (Guangdong University of Technology), Jianqi Liu (Guangdong University of Technology)
The 34th British Machine Vision Conference
Abstract
Few-shot classification and segmentation model realizes classification and segmentation by learning the feature correlation between a small number of samples. The lack of correlation learning between samples and the similarity of target foreground and background boundary pixels lead to segmentation errors, we propose Few-Shot Classification and Segmentation combining Bi-directional Feature Correlation and Boundary constraint(BFC-BL). Firstly, the correlation between query set and support set is calculated by cosine similarity to construct a 4D tensor. Then, a cross-scale bidirectional feature correlation fusion module (BFCP) is designed and embedded into the encoder structure to perform the interactive fusion of deep semantic correlation and shallow spatial correlation, while a bounding-constrained loss function is introduced to guide the model to learn the boundary information of the target foreground and background. Finally, a multi-level weight ratio loss function was constructed to make the network converge faster and generalize better. The experimental results show that compared with the ASNet method, the classification accuracy of the proposed method is increased by 1.7% and 1.8%, and the segmentation mean intersection over union ratio is increased by 1.3% and 1.4% on the Pascal-5^i. The code is publicly available at: https://github.com/XIAO1HAI/BFC-BL.Video
Citation
@inproceedings{Yang_2023_BMVC,
author = {Haibiao Yang and Zeng Bi and Pengfei Wei and Jianqi Liu},
title = {BFC-BL: Few-Shot Classification and Segmentation combining Bi-directional Feature Correlation and Boundary constraint},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0676.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by