CLODA: Cross Language Image Matching Based on Out-of-Distribution Data and Convolutional Block Attention Module for Weakly Supervised Semantic Segmentation

架構圖

1.準備數據集

PASCAL VOC2012

You will need to download the images (JPEG format) in PASCAL VOC2012 dataset at here and train_aug ground-truth can be found at here. Make sure your data/VOC2012 folder is structured as follows:

├── VOC2012/
|   ├── Annotations
|   ├── ImageSets
|   ├── SegmentationClass
|   ├── SegmentationClassAug
|   └── SegmentationObject

2.訓練W-OoD

cd ./CLODA/w-ood-main/

準備預訓練權重:

• Pre-trained model used in this paper: Download.
• Move it to "sess/"

準備分佈外數據:

• Download OoD images: Download
• Unzip and move it to "WOoD_dataset/openimages"
• Download a list of OoD images here and move 'ood_list.txt' to "WOoD_dataset/openimages"

訓練指令:

CUDA_VISIBLE_DEVICES=0 python run_sample.py --work_space="result/train_wood" --train_cam_pass True --make_cam_pass True --eval_cam_pass True  --cam_learning_rate 0.01 --voc12_root='../VOCdevkit/VOC2012/'

3.訓練CLODA

cd ./CLODA-master

將訓練好的wood權重放入./wood_pth

訓練指令:

CUDA_VISIBLE_DEVICES=0 python3 run_sample.py --voc12_root ../VOCdevkit/VOC2012/ --hyper 10,24,1,0.2 --clims_num_epoches 15 --cam_eval_thres 0.15 --work_space ./result/CLODA --cam_network net.resnet50_clims --train_clims_wood_clims_idea2_pass True --make_clims_pass True --eval_cam_pass True --clims_all True --z=0.1 

細化+生成偽標籤 指令:

CUDA_VISIBLE_DEVICES=0 python run_sample.py --voc12_root ../VOCdevkit/VOC2012/ --cam_eval_thres 0.15 --work_space ./result/CLODA --cam_network net.resnet50_clims --make_clims_pass True --cam_to_ir_label_pass True --train_irn_pass True --make_sem_seg_pass True --eval_sem_seg_pass True --clims_weights_name='res50_clims_best' --infer_list voc12/train_aug.txt --num_workers=1

4.偽標籤訓練DeepLabv2

將偽標籤放置好路徑../sem_seg/sem_seg_7533

更改指向config 路徑 :

  
--config-path configs/voc12_imagenet_pretrained_7533.yaml
--config-path configs/voc12_coco_pretrained_7533.yaml

運行:

CUDA_VISIBLE_DEVICES=0 bash run_voc12_coco_pretrained.sh
CUDA_VISIBLE_DEVICES=0 bash run_voc12_imagenet_pretrained.sh

Demo

https://youtu.be/3Iz0N3nJvZ4 https://youtu.be/xOGu4LfHNd8