MPN: Multi-task proposal network

doi:10.1016/j.knosys.2025.115032

MPN: Multi-task proposal network

2025-12-05

0

PRE

AI

DOI：10.1016/j.knosys.2025.115032

Original

Original request for help

Save

Abstract

En 中文

Multi-task learning (MTL) is a rapidly growing research area. However, existing MTL frameworks still face challenges in cross-task information sharing, collaborative optimization, and high computational overhead. To address these issues, we propose a Multi-task Proposal Network (MPN), which simultaneously handles two popular dense prediction tasks: depth estimation and semantic segmentation. MPN consists of three stages. The first stage performs feature extraction and information interaction, outputting a mask map with shared information and initial proposals. The second stage generates 1D proposals for each task; compared with directly decoding full-image features, task proposals incur lower overhead. The third stage fuses task proposals and the mask map, and completes prediction based on the respective label characteristics of depth estimation and semantic segmentation. To enhance multi-task optimization, we design an edge consistency loss that encourages the sharing of implicit spatial structures between tasks, thereby mitigating competition during training. Additionally, a semantic filtering strategy is designed to address the issue of speckle noise in semantic segmentation. Extensive experiments on the challenging Cityscapes and NYUD-v2 datasets for depth estimation and semantic segmentation demonstrate the superiority of the proposed method. Finally, we validate the zero-shot capability of MPN on SUN RGB-D dataset.

MPN: Multi-task proposal network

Abstract

Journal

Organization