Few-Shot Object Counting Using External  Visual Prompts

Brazhnyi, Anton

Home
→
Students Research & Project Works | Роботи студентів
→
Факультет прикладних наук
→
Освітня програма наук про дані
→
2024
→
View Item

dc.contributor.author	Brazhnyi, Anton
dc.date.accessioned	2024-08-22T09:58:06Z
dc.date.available	2024-08-22T09:58:06Z
dc.date.issued	2024
dc.identifier.citation	Brazhnyi Anton. Few-Shot Object Counting Using External Visual Prompts. Ukrainian Catholic University, Faculty of Applied Sciences, Department of Computer Sciences. Lviv 2024, 40 p.	uk
dc.identifier.uri	https://er.ucu.edu.ua/handle/1/4663
dc.language.iso	en	uk
dc.subject	Few-Shot Object Counting	uk
dc.subject	External Visual Prompts	uk
dc.title	Few-Shot Object Counting Using External Visual Prompts	uk
dc.type	Preprint	uk
dc.status	Публікується вперше	uk
dc.description.abstracten	Object counting is the task of estimating the number of specific objects present in an image. Similarly to other computer vision tasks, traditional object counting meth- ods typically require a large training dataset and are not suited for counting novel classes. Class-agnostic object counting, which is generally divided into few-shot and zero-shot approaches, aims to count arbitrary object categories. Few-shot count- ing requires manually labeled image patches depicting the object of interest, which is impractical in real-world applications. Zero-shot counting is primarily focused on using text prompts to specify the object without relying on manual annotations. However, text descriptions can be ambiguous and may not precisely convey ob- ject characteristics such as shape, texture, or size. Visual exemplars such as image patches act as a more direct reference, which leads to better generalizability and ac- curacy. In this work, we plan to explore the possibility of counting arbitrary objects in a few-shot manner without having humans in the loop. In particular, we are in- terested in utilizing a set of support images, which can be prepared in advance for a given object category and later used for all the query images. This would allow to accurately count specific objects without the need for extensive annotation.	uk