Contrastive Learning for Fine-grained Visual Recognition

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F49777513%3A23520%2F23%3A43969940" target="_blank" >RIV/49777513:23520/23:43969940 - isvavai.cz</a>

Výsledek na webu

<a href="http://svk.fav.zcu.cz/download/proceedings_svk_2023.pdf" target="_blank" >http://svk.fav.zcu.cz/download/proceedings_svk_2023.pdf</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Contrastive Learning for Fine-grained Visual Recognition

Popis výsledku v původním jazyce

Contrastive learning is a type of representation learning which retains a representation by comparing the input samples, e.g., images, video, text, and sound. Having good representation can be beneficial for the interpretability of Deep Neural Networks (DNNs) and for some downstream tasks like open-set recognition. Contrastive learning compares positive pairs of similar inputs and negative pairs of dissimilar inputs. The key component is the contrastive loss which measures the similarity between feature vectors and enforces minimization and maximization of the similarity between positive and negative pairs. Modern contrastive learning methods are often applied in self-supervised settings, while discriminative cross-entropy learning is widely used in supervised settings. In this work, we employ supervised contrastive learning to fine-tune DNNs for fine-grained recognition.

Název v anglickém jazyce

Contrastive Learning for Fine-grained Visual Recognition

Popis výsledku anglicky

Contrastive learning is a type of representation learning which retains a representation by comparing the input samples, e.g., images, video, text, and sound. Having good representation can be beneficial for the interpretability of Deep Neural Networks (DNNs) and for some downstream tasks like open-set recognition. Contrastive learning compares positive pairs of similar inputs and negative pairs of dissimilar inputs. The key component is the contrastive loss which measures the similarity between feature vectors and enforces minimization and maximization of the similarity between positive and negative pairs. Modern contrastive learning methods are often applied in self-supervised settings, while discriminative cross-entropy learning is widely used in supervised settings. In this work, we employ supervised contrastive learning to fine-tune DNNs for fine-grained recognition.

Druh

O - Ostatní výsledky

CEP obor

—

OECD FORD obor

20205 - Automation and control systems

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2023

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů