ECCV 2020 Daily - Thursday

3 Tony Ng 17 descriptor learning by co-author Yurun Tian . Local descriptors work only on a small patch, instead of a full large-size image. "The statistics between large images and small patches are quite different, so I thought it would be nice to also try it on global descriptors for large images and for landmarks,” Tony tells us. “What I found in this paper was that adding this second-order loss on top of my second- order attention gives the best results. It greatly improves upon the global methods and even matches some of the local methods that require 10 times the running speed, 10 times the running time , and three or four times the memory cost . This is a very promising result. I’m sure that the communities of image retrieval, descriptor learning, and visual localization will be very interested to hear about it.” and applicability are important, so I am focused on improving these global pooling methods.” One major problem Tony found with these methods is that they only calculate the first-order measure in terms of how to do the global pooling. Each feature in the feature map does not know about the other features and it is taking an average of the feature map to get the global descriptor. When he reviewed papers on self-attention and non- local blocks, he found it interesting that each feature could have the information of other features, without adding too much computational complexity. He incorporated this second-order attention in this work, alongside second-order loss, which was recently introduced in local DAILY T h u r s d a y