Jimmy1106/Spatial-NetVLAD

An encoding layer that can improve CNN performance on Image classification tasks.

Spatial-NetVLAD

• An encoding layer that can improve CNN performance on Image classification tasks.

Architecture

       

Performance comparisons

A Large Scale Fish Dataset

Pre-trained CNN	Pooling/Encoding layer	Accuracy
MoblieNet	Avg-pooling	93.3%
MoblieNet	NetVLAD	95.7%
MoblieNet	Spatial-NetVLAD	96.7%

Caltech101 Dataset

Pre-trained CNN	Pooling/Encoding layer	Accuracy
Vgg16	Avg-pooling	78.3%
Vgg16	NetVLAD	79.7%
Vgg16	Spatial-NetVLAD	86.9%

Reference works

• Madhan Chandrasekharan ,"Fish Classification-Transfer Learning Tensorflow" on Kaggle, 2021
• crlz182 , "Netvlad-Keras" on Github, 2019
• Yann Henon ,"keras-spp" on Github, 2017

Contact

• Email: jerming0515@gmail.com