Back to search

Institutional Repository of Vilnius University

Low resolution object segmentation /

Abstract

dc:description

Semantic segmentation of ultra low resolution (16×16 pixel) images is a challenging task because objects at low resolution do not have fine detail and thus ambigious boundaries. This problem is particularly relevant to privacy-preserving indoor robotics, where low-resolution sensing is deliberately used to prevent identification of people and documents. The state-of the art ULR2SS model addresses this with a joint super resolution and segmentation pipeline built on ESRGAN and DeepLabV3+, achieving competitive aggregate metrics while still producing imprecise object boundaries. This thesis investigates whether Active Boundary Loss (ABL) can improve boundary accuracy in the ULR2SS framework without degrading overall segmentation quality. Three fine tuning configurations were compared on a subset of the SUN RGBD indoor dataset: baseline fine tuning without ABL, ABL with a frozen super resolution generator (Frozen+ABL), and ABL with full joint training of both generator and segmentor (Full+ABL). In this single seed experimental setup, both ABL configurations improved all boundary metrics relative to the baseline. Full+ABL achieved the best boundary performance (Boundary F1 +15.3%, Hausdorff Distance −9.8%, Average Surface Distance −13.4%), while Frozen+ABL achieved the best observed pixel level metrics (mIoU 0.305, Pixel Accuracy 0.719); in this experimental setup, both ABL variants improved boundary metrics and did not reduce the observed pixel level metrics. These preliminary results indicate that ABL can improve boundary quality without changing the architecture. In this setup, freezing the super resolution generator gave the strongest pixel level scores, while joint training let the generator adapt toward boundary-preserving reconstructions. These trends need to be tested across multiple seeds and a larger training set before they can be treated as reliable.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Jogminas, Ignas,

Subjects

dc:subject × 1

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
eng

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Jogminas, Ignas,. Low resolution object segmentation /. Institutional Repository of Vilnius University, 2026. https://repository.vu.lt/VU:ELABAETD308337136&prefLang=en_US