Skip to content

Latest commit

 

History

History
68 lines (62 loc) · 2.21 KB

prepare_dataset.md

File metadata and controls

68 lines (62 loc) · 2.21 KB

Data preparation

Pascal VOC 2012

  • First download the Pascal VOC 2012 datasets use the scripts in the data dir.
cd data
sh download_and_convert_voc12.sh
  • Then download SBD annotations from here and download the Pascal VOC 2012 CAM_Label.

The folder structure is assumed to be:

WeakTr
├── data
│   ├── download_and_convert_voc12.sh
│   ├── voc12
│   │   ├── VOCdevkit
│   │   │   ├── VOC2012
│   │   │   │   ├── JPEGImages
│   │   │   │   ├── SegmentationClass
│   │   │   │   ├── SegmentationClassAug
│   │   │   │   ├── WeakTr_CAMlb_wCRF
├── voc12
│   ├── cls_labels.npy
│   ├── train_aug_id.txt
│   ├── train_id.txt
│   ├── val_id.txt

COCO 2014

  • First download the COCO 2014 datasets use the scripts in the data dir and download the COCO 2014 CAM_Label.
cd data
sh download_and_convert_coco.sh
cp ../coco/val_5000.txt coco/voc_format
cp ../coco/val_id.txt coco/voc_format/val.txt
cp ../coco/train_id.txt coco/voc_format/train.txt
  • Then download the COCO 2014 semantic segmentation labels from here

The folder structure is assumed to be:

WeakTr
├── data
│   ├── download_and_convert_coco.sh
│   ├── voc12
│   ├── coco
│   │   ├── images
│   │   ├── voc_format
│   │   │   ├── class_labels
│   │   │   ├── train.txt
│   │   │   ├── val.txt
│   │   │   ├── val_5000.txt
│   │   │   ├── WeakTr_CAMlb_wCRF_COCO
├── voc12
│   ├── cls_labels.npy
│   ├── train_aug_id.txt
│   ├── train_id.txt
│   ├── val_id.txt
├── coco
│   ├── cls_labels.npy
│   ├── train_id.txt
│   ├── train_1250_id.txt
│   ├── val_id.txt
│   ├── val_5000.txt