Reproduce “Hi-ResNet: Edge Detail Enhancement for High-Resolution Remote Sensing Segmentation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yuxia Chen, Pengcheng Fang, Jianhui Yu, Xiaoling Zhong, Xiaoming Zhang, Tianrui Li

Hi-ResNet, developed by researchers including those from Chengdu University of Technology and University of Southampton, enhances high-resolution remote sensing semantic segmentation by effectively addressing object scale variations and inter-class similarities through a multi-module architecture and a novel Class-agnostic Edge Aware (CEA) loss. The framework achieved an mIoU of 79.8% on Vaihingen and 52.6% on LoveDA, outperforming baselines while reducing parameters by over 50%.

Key insights

  • The Hi-ResNet base architecture, specifically with a deepened third stage and removed fourth stage, achieved superior mIoU with 30% fewer parameters than the HRNet baseline.
  • Unsupervised pre-training using MoCoV2 on large remote sensing datasets (Million-AID) significantly improved model performance by 3.6% mIoU compared to no pre-training.
  • The Class-agnostic Edge Aware (CEA) Loss effectively enhanced boundary perception and improved precision by 1% without adding parameters, demonstrating its ability to distinguish inter-class objects.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.