Ayan_kumar_bhunia a Deep One Shot Network for Query Based Logo Retrieval 2019

[TOC]

Summary of paper

Existing general purpose just cannot handle unseen new logos (not labelled logos)
in this work, they developed an easy-to-implement query based logo detection and localisation system by employing a one-shot learning technique using off-the-shelf neural network components.

Deep-learning based framework are largely data-driven, contrary to logo-dataset that have several image classes but few images.
need to be robust to new unseen logos, the model should be designed to satisfy the incremental demands for logo classes, contrary to existing methods which are limited to a set of seen logos and are not.

propose a scalable solution for the logo detection problem, they present a query-based logo search and detection system by employing a simple fully differentiable one-shot learning framework which can be used for new logo classes without further training the whole network.
to deal with the logos of varying sizes, we propose a novel one-shot framework through multi-scale conditioning that is specially designed to learn the similarity between the query image and target image at multiple scales and resolutions.

Conditioning module

the logo image is converted into a multichannel feature vector of unit spatial dimension (1 x 1 x 512)

Multi-scale conditioning

the Tile module helps to scale the 1x1x512 vector to target WxHx512
- tile: Constructs a tensor by repeating the elements of input