SOTAVerified

KGTN-ens: Few-Shot Image Classification with Knowledge Graph Ensembles

2022-11-06Code Available0· sign in to hype

Dominik Filipiak, Anna Fensel, Agata Filipowska

Code Available — Be the first to reproduce this paper.

Reproduce

Code

Abstract

We propose KGTN-ens, a framework extending the recent Knowledge Graph Transfer Network (KGTN) in order to incorporate multiple knowledge graph embeddings at a small cost. We evaluate it with different combinations of embeddings in a few-shot image classification task. We also construct a new knowledge source - Wikidata embeddings - and evaluate it with KGTN and KGTN-ens. Our approach outperforms KGTN in terms of the top-5 accuracy on the ImageNet-FS dataset for the majority of tested settings.

Tasks

Benchmark Results

DatasetModelMetricClaimedVerifiedStatus
ImageNet-FS (10-shot, all)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)83.46Unverified
ImageNet-FS (10-shot, novel)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)82.56Unverified
ImageNet-FS (1-shot, all)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)68.58Unverified
ImageNet-FS (1-shot, novel)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)62.73Unverified
ImageNet-FS (2-shot, all)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)75.45Unverified
ImageNet-FS (2-shot, novel)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)71.48Unverified
ImageNet-FS (5-shot, all)KGTN-ens (ResNet-50, h+g, max)Top-5 Accuracy (%)81.12Unverified
ImageNet-FS (5-shot, novel)KGTN-ens (ResNet-50, h+g, mean)Top-5 Accuracy (%)78.9Unverified

Reproductions