On the large-scale transferability of convolutional neural networks

Liang Zheng*, Yali Zhao, Shengjin Wang, Jingdong Wang, Yi Yang, Qi Tian

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Given the overwhelming performance of the Convolutional Neural Network (CNN) in the computer vision and machine learning community, this paper aims at investigating the effective transfer of the CNN descriptors in generic and fine-grained classification at a large scale. Our contribution consists in providing some simple yet effective methods in constructing a competitive baseline recognition system. Comprehensively, we study two facts in CNN transfer. (1) We demonstrate the advantage of using images with a properly large size as input to CNN instead of the conventionally resized one. (2) We benchmark the performance of different CNN layers improved by average/max pooling on the feature maps. Our evaluation and observation confirm that the Conv5 descriptor yields very competitive accuracy under such a pooling strategy. Following these good practices, we are capable of producing improved performance on seven image classification benchmarks.

Original languageEnglish
Title of host publicationTrends and Applications in Knowledge Discovery and Data Mining - PAKDD 2018 Workshops, BDASC, BDM, ML4Cyber, PAISI, DaMEMO, Revised Selected Papers
EditorsMohadeseh Ganji, Lida Rashidi, Benjamin C.M. Fung, Can Wang
PublisherSpringer Verlag
Pages27-39
Number of pages13
ISBN (Print)9783030045029
DOIs
Publication statusPublished - 2018
Externally publishedYes
Event22nd Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2018 - Melbourne, Australia
Duration: 3 Jun 20183 Jun 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11154 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference22nd Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2018
Country/TerritoryAustralia
CityMelbourne
Period3/06/183/06/18

Fingerprint

Dive into the research topics of 'On the large-scale transferability of convolutional neural networks'. Together they form a unique fingerprint.

Cite this