Skip to content

Latest commit

 

History

History
59 lines (56 loc) · 3.5 KB

DATASET.md

File metadata and controls

59 lines (56 loc) · 3.5 KB
Raw

Datasets

We use the following datasets in our experiments:

Dataset structure

For simplicity, our datasets are structured in the following way:

/CIM/data/
├── VOC2012/
│   ├── annotations/
│   ├── JPEGImages/
│   ├── COB_SBD_trainaug/
│   ├── COB_SBD_val/
│   └── Center_points/
│
├── coco2017/
│   ├── annotations/
│   ├── train2017/
│   ├── val2017/
│   ├── test2017/
│   └── COB-COCO/
│
├── model_weight/
│   ├── prm_voc.pth
│   ├── prm_coco.pth
│   ├── vgg16_caffe.pth
│   └── hrnetv2_w48_imagenet_pretrained.pth
│ 
├── label_assign/
│   ├── voc_2012_label_assign.pkl
│   └── coco_2017_label_assign.pkl
│
├── cob/
│   ├── voc_2012_trainaug.pkl
│   ├── voc_2012_val.pkl
│   ├── coco_2017_train.pkl
│   ├── coco_2017_val.pkl
│   └── coco_2017_test.pkl
│
├── cob_asy_iou/
│   ├── VOC2012/
│   └── coco2017/
│
└── cob_iou/
    ├── VOC2012/
    └── coco2017/

Note:

  • VOC2012/annotations/ is a folder containing label files in json format. You can convert .xml format annotations to coco format, or you can use our pre-processed annotations here.
  • VOC2012/COB_SBD_trainaug/, VOC2012/COB_SBD_val/ and coco2017/COB-COCO/ are folders containing COB files. You can download VOC2012/COB_SBD_trainaug/, VOC2012/COB_SBD_val/ from here.coco2017/COB-COCO/ can be downloaded from here.
  • VOC2012/Center_points/ is a folder containing point-level label files. We download them from BESTIE.
  • model_weight/ is a folder containing weight of models. Weights can be downloaded from here. hrnetv2_w48_imagenet_pretrained.pth can be downloaded from here.
  • label_assign/ contains pre-computed pseudo labels for VOC2012 and COCO2017 datasets, link is here. It also can be created by running python tools/pre/AGPL_label_assign.py.
  • cob/ is a folder containing two dataset proposals. The pkl files contain proposals that are scaled to a size of 7*7. These files will be used in RoiAlign operation. They can be downloaded from here. They also can be created by running python tools/pre/generate_7_7_voc.py and python tools/pre/generate_7_7_coco.py.
  • cob_iou/ can be downloaded from here. It also can be created by running python tools/pre/create_cob_iou.py.
  • cob_asy_iou/ can be downloaded from here. It also can be created by running python tools/pre/create_cob_asy_iou.py.