语义分割模型

Segment-Anything(SAM)

SAM2

Grounding-DINO

Grounded-SAM

Geospatial-SAM

Samba

  1. 安装
pip install causal-conv1d==1.1.1 
pip install mamba-ssm
pip install apex-amp, pyyaml, timm, tlt （apex-amp没找到，没装）
  1. 数据集 loveDA

​ 下载的Train.zip,VAL.zip用于训练生成权重文件，自行下载解压后，放置在 samba-path/data/loveDA/下，并按照以下结构调整:

ann_dir下存放Train和VAL的mask_pngs文件内图片，即瓦片的掩码

img_dir 下存放Train和VAL的瓦片文件

  • ann_dir
    • train
      • .png
    • val
      • .png
  • img_dir
    • train
      • .png
    • val
      • .png
3. 配置文件
  • The model file Samba.py can be found in samba-path/mmseg/models/backbones/
  • The config file samba_upernet.py for the combination of backbone and decoder head samba_upernet can be found in /samba-path/configs/base/models
  • The config file samba_upernet-15k_loveda-512x512_6e4.py for training can be found in /samba-path/configs/samba/
4. 训练 Samba

双卡4090训练速度大概在2个小时，针对Train的城镇数据

bash tools/dist_train.sh /samba-path/configs/samba/samba_upernet-15k_loveda-512x512_6e4.py 2 --work-dir /samba-path/output/sambaupernet
bash tools/dist_train.sh /samba-path/configs/samba/samba_upernet-15k_potsdam-512x512_6e4.py 2 --work-dir /samba-path/output/sambaupernet
bash tools/dist_train.sh /samba-path/configs/samba/samba_upernet-15k_vaihingen-512x512_6e4.py 2 --work-dir /samba-path/output/sambaupernet
测试 Samba

有点没搞明白，测试数据也需要提供masks_png,不然就报错了······

bash tools/dist_test.sh /samba-path/configs/samba/samba_upernet-15k_loveda-512x512_6e4.py \ /samba-path/output/sambaupernet/iter_15000.pth 2 --out /mmsegmentation/visout/sambaupernet
bash tools/dist_test.sh /samba-path/configs/samba/samba_upernet-15k_potsdam-512x512_6e4.py \ /samba-path/output/sambaupernet/iter_15000.pth 2 --out /mmsegmentation/visout/sambaupernet
bash tools/dist_test.sh /samba-path/configs/samba/samba_upernet-15k_vaihingen-512x512_6e4.py \ /samba-path/output/sambaupernet/iter_15000.pth 2 --out /mmsegmentation/visout/sambaupernet
github搜索关键词

https://github.com/liliu-avril/Awesome-Segment-Anything

https://github.com/Hedlen/awesome-segment-anything