Multi-Scale Semantic Enrichment and Dual Angular Margin Contrast for Few-Shot Class Incremental Learning


Riya Verma (Indian Institute of Technology, Madras), Sukhendu Das (Indian Institute of Technology Madras)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Few-shot class incremental learning (FSCIL) faces the dual challenge of learning new classes from limited data without forgetting previously learned classes. The predominant framework addressing FSCIL relies on extracting a single feature vector per image, resulting in inadequate generalization to new classes with limited samples and misclassifying these novel classes into existing base classes or similar novel classes. To address this, we propose the Multi-Scale Semantic Enrichment method, which extracts several feature vectors from each image and captures rich semantic details across different scales and levels of abstraction. Our Dual Angular Margin Contrast (DAMC) prototype learning framework ensures compact intra-class embeddings and enhances inter-class separability, preserving space for novel classes and encouraging the model to learn features representative of the class. We employ a non-parametric self-attention mechanism to prioritize the most informative samples to obtain weighted prototypes. Additionally, we apply Layerwise Feature Augmentation for a more complex and diverse feature landscape and leverage set-based distance metrics for refined inference. The performance of our proposed method, when verified over CIFAR100, CUB200, and miniImageNet datasets, sets new state-of-the-art.

Citation

@inproceedings{Verma_2024_BMVC,
author    = {Riya Verma and Sukhendu Das},
title     = {Multi-Scale Semantic Enrichment and Dual Angular Margin Contrast for Few-Shot Class Incremental Learning},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0828.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection