Pretrained_Dataset.md

:notebook_with_decorative_cover: Prepare Dataset

---

:racehorse:COCO

1. Download COCO

   you can also modify the $DATA_ROOT here

   bash get_COCO.sh
   

2. Since COCO dataset does not store images in corresponding class directory, we have to add a fake root directory under train_2014

   cd $DATA_ROOT
   mkdir -p tmp_root
   mv train2014 tmp_root/train2014
   

    DATA_ROOT(For COCO)
    ├── tmp_root
    │   ├── train2014
    │   │   ├── COCO_train2014_XXXXXXXXXX80.jpg.png
    │   │   ├── COCO_train2014_XXXXXXXXXX81.jpg.png   
    ├── val2014
    │   ├── COCO_val2014_XXXXXXXXXX01
    │   ├── COCO_val2014_XXXXXXXXXX02.png
    │   ├── COCO_val2014_XXXXXXXXXX03.png   
   

3. If the pre-trained framework utilizes the annotation .txt for training data (Like InsLoc), create it with this file

---

:dog:ImageNet

Download Imagenet

1. Obtain the following files at Image-net.org

   • ILSVRC2012_img_train.tar
   • ILSVRC2012_img_val.tar
   • ILSVRC2012_devkit_t12.tar.gz

2. Preprocess ImageNet validation set:

   # Modify the DATA_ROOT in shell script
   bash imagenet_preprocess.sh ${DATA_ROOT}
   

3. The directory structure should look like this:

   DATA_ROOT(For ImageNet)
   ├── train
   │   ├── nXXXXX
   │   │   ├── nXXXXX_01.png
   │   │   ├── nXXXXX_02.png   
   ├── val
   │   ├── nXXXXX
   │   │   ├── nXXXXX_01.png
   │   │   ├── nXXXXX_02.png