Masofish / image_seg

A Strong Baseline for Image Semantic Segmentation

A Strong Baseline for Image Semantic Segmentation

Introduction

This project is an open source semantic segmentation toolbox based on PyTorch. It is based on the codes of our Tianchi competition in 2021 (https://tianchi.aliyun.com/competition/entrance/531860/introduction).
In the competition, our team won the third place (please see Tianchi_README.md).

Overview

The master branch works with PyTorch 1.6+.The project now supports popular and contemporary semantic segmentation frameworks, e.g. UNet, DeepLabV3+, HR-Net etc.

Requirements

Support

Backbone

• ResNet (CVPR'2016)
• SeNet (CVPR'2018)
• IBN-Net (CVPR'2018)
• EfficientNet (CVPR'2020)

Methods

• UNet
• DLink-Net
• Res-UNet
• Efficient-UNet
• Deeplab v3+
• HR-Net

Tricks

• MixUp /CutMix /CopyPaste
• SWA
• LovaszSoftmax Loss /LargeMarginSoftmax Loss
• FP16
• Multi-scale

Tools

• large image inference (cut and merge)
• post process (crf/superpixels)

Quick Start

Train a model

python train.py --config_file ${CONFIG_FILE} 

• CONFIG_FILE: File of training config about model

Examples:
We trained our model in Tianchi competition according to the following script:
Stage 1 (160e)

python train.py --config_file configs/tc_seg/tc_seg_res_unet_r34_ibn_a_160e.yml

Stage 2 (swa 24e)

python train.py --config_file configs/tc_seg/tc_seg_res_unet_r34_ibn_a_swa.yml

Inference with pretrained models

python inference.py --config_file ${CONFIG_FILE} 

• CONFIG_FILE: File of inference config about model

Predict large image with pretrained models

python predict_demo.py --config_file ${CONFIG_FILE} --rs_img_file ${IMAGE_FILE_PATH} --temp_img_save_path ${TEMP_CUT_PATH} -temp_seg_map_save_path ${TEMP_SAVE_PATH} --save_seg_map_file ${SAVE_SEG_FILE} 

• CONFIG_FILE: File of inference config about model
• IMAGE_FILE_PATH: File of large input image to predict
• TEMP_CUT_PATH: Temp folder of small cutting samples
• TEMP_SAVE_PATH: Temp folder of predict results of cutting samples
• SAVE_SEG_FILE: Predict result of the large image