EXPERIMENTAL STUDY OF DISTANCE MAP DECODER ARCHITECTURAL CONFIGURATIONS FOR INSTANCE SEGMENTATION BY TEXT QUERY

Read full text Summarise with VIRAT
Authors: Andrii Romanovych Kovtunenko, Sergii Mashtalir
Cited by: 0
References: 15

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This article presents an experimental study of the architectural configurations of distance map decoder in the InstanceCLIPSeg model for instance segmentation by text query. We in-vestigate the influence of various mechanisms for restoring spatial resolution (bilinear inter-polation, PixelShuffle), the use of coordinate convolutions (CoordConv), and multi-level feature fusion strategies. Based on the evaluation of nine configurations on the LVIS and PhraseCut datasets, it was found that a hybrid architecture with PixelShuffle and single-stage feature fusion from transformer layers achieves the best results (mean Dice 0.2374), outper-forming baseline approaches. The redundancy of coordinate channels in the presence of mul-ti-level fusion was revealed.

Fields of Study & Subjects

Advanced Neural Network Applications Multimodal Machine Learning Applications Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
System technologies
Volume
2
Issue
163
Pages
3–12
Publication date
31 March 2026
Language
EN