A lightweight depthwise separable Dilated Multires Network (DDMnet) for instance segmentation

Nikolaos Detsikas

doi:10.1007/s11042-026-21907-1

A lightweight depthwise separable Dilated Multires Network (DDMnet) for instance segmentation

2026-09-14

0

PRE

AI

N

Nikolaos Detsikas

C

Christos Chatzisavvas

N

Nikolaos Mitianoudis *

I

Ioannis Pratikakis

DOI：10.1007/s11042-026-21907-1

Original

Original request for help

Save

Abstract

En 中文

The differentiation and identification of individual objects within a scene is a fundamental problem in modern computer vision. This task, known as instance segmentation, has received extensive attention in recent years due to the advances in Deep Learning. Despite notable progress, many state-of-the-art approaches rely on increasingly complex architectures with high computational requirements, making them unsuitable for real-time applications, such as autonomous driving. In this work, we propose a compact model for addressing the instance segmentation problem, that integrates a novel convolutional encoder, named DDMnet, with the Mask2Former segmentation framework. The proposed network achieves low computational complexity and delivers performance close to state-of-the-art methods on benchmark datasets, including Cityscapes and ADE20K, demonstrating its potential for real-time deployment.

Keywords:

Cityscapes

ADE20K

Instance segmentation

Deep learning

A lightweight depthwise separable Dilated Multires Network (DDMnet) for instance segmentation

Abstract

Journal

Organization