Detekcija objekata nad otvorenim rječnikom

master's thesis

• Title (english): Open-vocabulary object detection
• Author: Bruno Ćorić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Committee member: Adrian Satja Kurdija (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-10, Croatia

Abstract

Detekcija objekata nad otvorenim vokabularom otvara nova područja za korištenje računalnog vida. Klasični pristupi za detekciju objekata se sastoje od treniranja modela modela računalnog vida nad zatvorenim skupom kategorija te takvi modeli nisu fleksibilni što se tiče korištenja u nove svrhe. Ovakvim pristupom broj kategorija koje model može detektirati je beskonačan te model ne mora u fazi treninga biti učen na određenoj kategoriji objekata kako bi ih znao prepoznati. U ovom radu trenirali smo model nad samostalno napravljenim skupom podataka koji sadrži slike s bespilotnih letjelica. Prikazani su rezultati modela nad skupom podataka VisDrone2019 u usporedbi sa drugim modelima koji su samo trenirani kako bi radili nad VisDrone2019 skupu podataka. Također je prikazano korištenje ovakvih modela nad deskriptivnom kategorijama objekata te prilagođavanje ovako treniranog modela potpuno drugim domenama.

Abstract (english)

Object detection with open vocabulary opens new areas for the use of computer vision. Classical approaches to object detection involve training computer vision models on a closed set of categories, and such models are not flexible for new purposes. With this approach, the number of categories that the model can detect is infinite, and the model does not need to be trained on a specific category of objects during the training phase to be able to recognize them. In this paper, we trained a model on a custom-made dataset containing images from unmanned aerial vehicles. The results of the model on the VisDrone2019 dataset are presented and compared with other models that were only trained to work on the VisDrone2019 dataset. The use of such models with descriptive categories of objects is also demonstrated, as well as the adaptation of a model to completely different domains.

• Keywords: transformer; detekcija objekata; bespilotne letjelice
• Keywords (english): transformer; object detection; unmanned aerial vehicles
• Language: croatian
• URN:NBN: urn:nbn:hr:168:310168
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 20:22:18