Rethinking Fisher Vectors for Second-order Multi-scale Deep Feature Aggregation in Scalable Scene Retrieval

概要

Aggregating CNN responses into a compact discriminative vector, given their lack of geometric invariance, is a challenge. To address this challenge, we propose a codebook-free method to encode second-order information of multi-scale CNN features within the Fisher Kernel framework. We use a scalable retrieval pipeline to show the efficacy of the proposed vector. The increased discriminative power is confirmed by experimental evaluation of our method on two difficult benchmarks (INRIA Holidays, UKBench). The significant improvement in retrieval accuracy over state-of-the-art approaches comes with only a slight dimension increase.

参考リンク

• Paper ↗

リンク先のライセンス・再現性・データ条件をご確認ください。自動分類や実装の目安は保証ではありません。

実装を始めるプロンプト

参考リンクを確認したうえで、次の内容をコピーして利用できます。

# 論文・記事
Rethinking Fisher Vectors for Second-order Multi-scale Deep Feature Aggregation in Scalable Scene Retrieval

# 概要
Aggregating CNN responses into a compact discriminative vector, given their lack of geometric invariance, is a challenge. To address this challenge, we propose a codebook-free method to encode…

# 原典
https://arxiv.org/abs/2610.31146

上記の手法を実装してください。実行前に依存関係とライセンスを確認してください。