pipette
ESEspañol

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

Alexy Bhowmick, Sarat Saharia, Shyamanta M. Hazarika

Preprint

In the authors' words

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.

Main resultLimitation the authors admit

Appeared: Friday, October 23. arXiv. Preprint, not yet peer-reviewed.

Authors' comment: 8 pages, 4 figures