CPNet: Accurate Contour Preservation for Semantic Segmentation by Mitigating the Impact of Pseudo-boundaries

This repository contains the implementation details of our paper "Accurate Contour Preservation for Semantic Segmentation by Mitigating the Impact of Pseudo-boundaries".

Source Files

• configs/* : config files
• mmseg/evaluation/metrics/boundary_metric.py: the calculation of boundary IoU
• mmseg/models/decode_heads/*, mmseg/models/necks/* , mmseg/models/segmentors/cascade_encoder_decoder_3.py: model files
• mmse/models/losses/*: loss function

Environment Requirements

• Linux/MacOS/Windows
• Python 3.7+
• CUDA 10.2+
• PyTorch 1.8+
• mmcv 2.0
• mmsegmentation

Noted: The code is based on mmsegmentation v1.1.2. Just follow this guidance get_started to install the environment and copy or update the code files of this repository to the corresponding files, and everything will be ready.

Dataset

• Vaihingen
• Potsdam