prikaz prve stranice dokumenta Open-set segmentation of images by means of negative examples
  Download
PDF 40.41 MB
doctoral thesis
Open-set segmentation of images by means of negative examples
Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing, 2023. urn:nbn:hr:168:468120
Bevandić, Petra
University of Zagreb
Faculty of Electrical Engineering and Computing

Institutional repository: FER Repository

Cite this document

Bevandić, P. (2023). Open-set segmentation of images by means of negative examples (Doctoral thesis). Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing. Retrieved from https://urn.nsk.hr/urn:nbn:hr:168:468120

Bevandić, Petra. "Open-set segmentation of images by means of negative examples." Doctoral thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2023. https://urn.nsk.hr/urn:nbn:hr:168:468120

Bevandić, Petra. "Open-set segmentation of images by means of negative examples." Doctoral thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2023. https://urn.nsk.hr/urn:nbn:hr:168:468120

Bevandić, P. (2023). 'Open-set segmentation of images by means of negative examples', Doctoral thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, accessed 02 June 2024, https://urn.nsk.hr/urn:nbn:hr:168:468120

Bevandić P. Open-set segmentation of images by means of negative examples [Doctoral thesis]. Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing; 2023 [cited 2024 June 02] Available at: https://urn.nsk.hr/urn:nbn:hr:168:468120

P. Bevandić, "Open-set segmentation of images by means of negative examples", Doctoral thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, 2023. Available at: https://urn.nsk.hr/urn:nbn:hr:168:468120

Metadata
TitleOpen-set segmentation of images by means of negative examples
Title (croatian)Segmentacija slika nad otvorenim skupom razreda putem negativnih primjera
AuthorPetra Bevandić
MentorSiniša Šegvić (mentor)
Committee memberZoran Kalafatić (predsjednik povjerenstva)
Committee memberMarko Subašić (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb
Defense date and country2023-07-12, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Universal decimal classification
(UDC)		004 - Computer science and technology. Computing. Data processing
Abstract
Open-set recognition simulates real-world settings by introducing outlier samples into model evaluation. This problem is more complex in the dense prediction setting because input can contain a mix of in-distribution and out-of-distribution segments. Open-set recognition tasks can be approached through simultaneous classification and anomaly detection. This work investigates methods for open-set prediction which are trained by means of negative examples. We assume that we train domain-specific models and can thus sample negatives from large and diverse datasets (e.g. ImageNet). These types of negatives introduce additional noise into training because they may also contain some positive content. This thesis will investigate various facets of training on noisy negative samples: from choosing the best architecture and loss to different methods of combining positive and negative samples during batch formation. The presented experiments validate our contributions on standard open-set recognition benchmarks.
Abstract (croatian)
Prepoznavanje nad otvorenim skupom razreda simulira rad modela u stvarnim uvjetima uvođenjem izvandistribucijskih primjera u evaluaciju modela. U gustoj predikciji postoji i dodatna razina složenosti jer ulazni primjeri mogu biti mješavina unutardistribucijskih i izvandistribucijskih segmenata. Problem prepoznavanja nad otvorenim skupom razreda možemo rješavati modelima koji istovremeno provode klasifikaciju i detekciju anomalija. U ovome ćemo radu istražiti metode za gustu predikciju nad otvorenim skupom razreda uključivanjem negativnih primjera u postupak učenja. Pretpostavljamo da su modeli koje učimo specijalizirani zbog čega se kao negativi mogu koristiti primjeri iz velikih, raznolikih skupova slika (npr. ImageNet). Takvi negativi unose šum u treniranje s obzirom da mogu sadržavati i unutardistribucijske uzorke. Disertacija će proučiti postupak uključivanja šumovitih negativa u postupak treniranja - od odabira arhitekture modela i gubitka do kombiniranja pozitivnih i negativnih primjera prilikom učenja. Doprinosi će biti vrednovani na standardnim skupovima za validaciju guste detekcije izvandistribucijskih primjera.
Keywords
Keywords (croatian)
Languageenglish
URN:NBNurn:nbn:hr:168:468120
Promotion2024
Study programmeTitle: Computer Science
Study programme type: university
Study level: postgraduate
Academic / professional title: Doktor znanosti (Doktor znanosti)
Type of resourceText
Extentxi, 82 str. : graf. prikazi.
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2024-02-21 14:51:28