Skip to content

Latest commit

 

History

History
401 lines (320 loc) · 12.5 KB

README.md

File metadata and controls

401 lines (320 loc) · 12.5 KB
Raw

Detectron2 Builtin Datasets

Detectron2 has builtin support for a few datasets.

The datasets are assumed to exist in a directory specified by the environment variable DETECTRON2_DATASETS.

Under this directory, following here to prepare COCO, LVIS, cityscapes, Pascal VOC and ADE20k.

The expected structure is described below.

$DETECTRON2_DATASETS/
  coco/
  lvis/
  cityscapes/
  VOC20{07,10,12}/
  ADEChallengeData2016/

You can set the location for builtin datasets by export DETECTRON2_DATASETS=/path/to/datasets. If left unset, the default is ./datasets relative to your current working directory.

APE Builtin Datasets

Expected dataset structure for COCO and LVIS

$DETECTRON2_DATASETS/
  coco/
    annotations/
      instances_{train,val}2017.json
      panoptic_{train,val}2017.json
    {train,val}2017/
    panoptic_{train,val}2017/
    panoptic_stuff_{train,val}2017/
    panoptic_semseg_{train,val}2017/
  lvis/
    lvis_v1_{train,val}.json
    lvis_v1_{train,val}+coco_mask.json
    lvis_v1_{train,val}+coco_mask_cat_info.json

panoptic_semseg_{train,val}2017/ are generated by runing

python3 datasets/prepare_coco_semantic_annos_from_panoptic_annos.py

lvis_v1_{train,val}+coco_mask.json are generated by running

python3 datasets/tools/lvis/merge_lvis_coco.py

lvis_v1_{train,val}+coco_mask_cat_info.json are generated by running

python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/lvis/lvis_v1_train+coco_mask.json
python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/lvis/lvis_v1_val+coco_mask.json

Expected dataset structure for Objects365:

$DETECTRON2_DATASETS/
  objects365/
    annotations/
      zhiyuan_objv2_{train,val}.json
      objects365_{train,val,minival}_fixname.json
    train/
      images/
    val/
      images/

objects365_train_fixname.json and objects365_val_fixname.json are generated by running

python3 datasets/tools/objects3652coco/get_image_info.py --image_dir datasets/objects365/train/ --json_path datasets/objects365/annotations/zhiyuan_objv2_train.json --output_path datasets/objects365/annotations/image_info_train.txt
python3 datasets/tools/objects3652coco/get_image_info.py --image_dir datasets/objects365/val/ --json_path datasets/objects365/annotations/zhiyuan_objv2_val.json --output_path datasets/objects365/annotations/image_info_val.txt

python3 datasets/tools/objects3652coco/convert_annotations.py --root_dir datasets/objects365/ --image_info_path datasets/objects365/annotations/image_info_train.txt --subsets train --apply_exif
python3 datasets/tools/objects3652coco/convert_annotations.py --root_dir datasets/objects365/ --image_info_path datasets/objects365/annotations/image_info_val.txt --subsets val --apply_exif
python3 datasets/tools/objects3652coco/convert_annotations.py --root_dir datasets/objects365/ --image_info_path datasets/objects365/annotations/image_info_val.txt --subsets minival --apply_exif

python3 datasets/tools/objects3652coco/fix_o365_names.py --ann datasets/objects365/annotations/objects365_train.json
python3 datasets/tools/objects3652coco/fix_o365_names.py --ann datasets/objects365/annotations/objects365_val.json
python3 datasets/tools/objects3652coco/fix_o365_names.py --ann datasets/objects365/annotations/objects365_minival.json

As Objects365 is large, we generate annotation file for each image separetely

python3 datasets/tools/generate_img_ann_pair.py --json_path datasets/objects365/annotations/objects365_train_fixname.json --image_root datasets/objects365/train/

Expected dataset structure for OpenImages:

$DETECTRON2_DATASETS/
  openimages/
    annotations/
      openimages_v6_{train,val}_bbox.json
      openimages_v6_{train,val}_bbox_nogroup.json
      openimages_v6_{train,val}_bbox_cat_info.json
      openimages_v6_{train,val}_bbox_nogroup_cat_info.json
    train/
    validation/

openimages_v6_{train,val}_bbox.json are generated by running

python3 datasets/tools/openimages2coco/convert_annotations.py --path datasets/openimages/ --version v6 --subset train --task bbox --apply-exif
python3 datasets/tools/openimages2coco/convert_annotations.py --path datasets/openimages/ --version v6 --subset val --task bbox --apply-exif

openimages_v6_{train,val}_bbox_nogroup.json are generated by running

python3 datasets/tools/openimages2coco/convert_annotations.py --path datasets/openimages/ --version v6 --subset train --task bbox --apply-exif --exclude-group
python3 datasets/tools/openimages2coco/convert_annotations.py --path datasets/openimages/ --version v6 --subset val --task bbox --apply-exif --exclude-group

*_cat_info.json are generated by running

python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/openimages/annotations/openimages_v6_train_bbox.json
python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/openimages/annotations/openimages_v6_val_bbox.json
python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/openimages/annotations/openimages_v6_train_bbox_nogroup.json
python3 datasets/tools/lvis/add_category_info_frequence.py --json_path datasets/openimages/annotations/openimages_v6_val_bbox_nogroup.json

Finally, runing

python3 datasets/tools/generate_img_ann_pair.py --json_path datasets/openimages/annotations/openimages_v6_train_bbox.json --image_root datasets/openimages/train/

Expected dataset structure for VisualGenome:

$DETECTRON2_DATASETS/
  visualgenome/
    annotations/
      visualgenome_77962_box.json
      visualgenome_77962_box_{train,val}.json
      visualgenome_region.json
      visualgenome_region_{train,val}.json
      visualgenome_77962_box_and_region.json
      visualgenome_77962_box_and_region__{train,val}.json
    VG_100K/
    VG_100K_2/

visualgenome_*.json are generated by running

python3 datasets/tools/visualgenome2coco/convert_annotations_object.py -p datasets/visualgenome/ --apply-exif --object_list "" --num_objects 99999999 --min_box_area_frac 0.0

python3 datasets/tools/visualgenome2coco/convert_annotations_region.py -p datasets/visualgenome/ --apply-exif --object_list "" --num_objects 99999999 --min_box_area_frac 0.0

Expected dataset structure for SA-1B:

$DETECTRON2_DATASETS/
  SA-1B/
    images/
    sam1b_instance_1000000.json
    ...
    sam1b_instance.json

sam1b_instance*.json are generated by running

python tools/sa1b2coco/image+json.py --image_root datasets/SA-1B/images/ --json_path datasets/SA-1B/sam1b_instance

Expected dataset structure for RefCOCO:

$DETECTRON2_DATASETS/
  SeqTR/
    mixed/
    refcocog-google/
        instances_cocofied_{train,val}.json
    refcocog-umd/
        instances_cocofied_{train,val,test}.json
    refcoco-unc/
        instances_cocofied_{train,val,testA,testB}.json
    refcocoplus-unc/
        instances_cocofied_{train,val,testA,testB}.json
    refcoco-mixed/
        instances_cocofied_train.json
    refcoco-mixed_group-by-image/
        instances_cocofied_train.json

Download the preprocessed json files from here

refcoco-mixed/ and some instances_cocofied_*.json are generated by running

python3 datasets/tools/seqtr2coco/convert_mix_ref.py

refcoco-mixed_group-by-image// and its instances_cocofied_train.json are generated by running

python3 datasets/tools/seqtr2coco/convert_refcoco_mixed_group_by_image.py

Expected dataset structure for GQA:

$DETECTRON2_DATASETS/
  gqa/
    images/
    gqa_region_{train,val}.json
    gqa_region.json

gqa_region*.json are generated by running

python3 datasets/tools/gqa2coco/convert.py --data_path datasets/gqa/ --img_path datasets/gqa/images --sg_path datasets/gqa/ --vg_img_data_path datasets/visualgenome/annotations/ --out_path datasets/gqa/

Expected dataset structure for PhraseCut:

$DETECTRON2_DATASETS/
  phrasecut/
    images/
    phrasecut_{train,val,miniv,test}.json

phrasecut_*.json are generated by running

python3 datasets/tools/phrasecut2coco/convert.py --data_path datasets/phrasecut/ --img_path datasets/phrasecut/images --out_path datasets/phrasecut/

Expected dataset structure for Flickr30k:

$DETECTRON2_DATASETS/
  flickr30k/
    flickr30k-images/
    flickr30k_separateGT_{train,val.test}.json

flickr30k_separateGT_*.json are generated by running

python3 datasets/tools/flickr2coco/convert.py --flickr_path datasets/flickr30k/flickr30k_entities/ --out_path datasets/flickr30k/

Expected dataset structure for ODinW:

$DETECTRON2_DATASETS/
  odinw/
    AerialMaritimeDrone/
    AmericanSignLanguageLetters/
    ...
    WildfireSmoke/

After download, update json files by runing

python3 datasets/tools/odinw/convert.py

This is because

https://github.com/cocodataset/cocoapi/issues/507#issuecomment-857272753

Expected dataset structure for SegInW:

$DETECTRON2_DATASETS/
  seginw/
    Airplane-Parts/
    Bottles/
    ...
    Watermelon/

Expected dataset structure for Roboflow100:

$DETECTRON2_DATASETS/
  rf100/
    4-fold-defect/
    abdomen-mri/
    ...
    x-ray-rheumatology/

Expected dataset structure for ADE20k-Full:

ADE20K_2021_17_01/
  images/
  images_detectron2/
  annotations_detectron2/
  index_ade20k.pkl
  objects.txt

The directories images_detectron2 and annotations_detectron2 are generated by running

python datasets/prepare_ade20k_full_sem_seg.py

Expected dataset structure for BDD10k:

$DETECTRON2_DATASETS/
  bdd100k/
    images/
    labels/
      pan_seg/
        coco_pano/
        meta/
        ...
      ...
    seg/

coco_pano and meta is generated by running

wget https://github.com/shenyunhang/APE/releases/download/0/bdd_generated.tar.gz
tar xvzf bdd_generated.tar.gz

Expected dataset structure for PC459 and PC59:

$DETECTRON2_DATASETS/
  VOCdevkit/
    VOC2010/
      Annotations/
      ImageSets/
      JPEGImages/
      SegmentationClass/
      SegmentationObject/
      # below are from https://www.cs.stanford.edu/~roozbeh/pascal-context/trainval.tar.gz
      trainval/
      labels.txt
      59_labels.txt # https://www.cs.stanford.edu/~roozbeh/pascal-context/59_labels.txt
      pascalcontext_val.txt # https://drive.google.com/file/d/1BCbiOKtLvozjVnlTJX51koIveUZHCcUh/view?usp=sharing
      # below are generated
      annotations_detectron2/
        pc459_val/
        pc59_val

It starts with a tar file VOCtrainval_03-May-2010.tar. You may want to download the 5K validation set here.

The directory annotations_detectron2 is generated by running

python datasets/prepare_pascal_context.py

Expected dataset structure for VOC:

$DETECTRON2_DATASETS/
  VOCdevkit/
    VOC2012/
      Annotations/
      ImageSets/
      JPEGImages/
      SegmentationClass/
      SegmentationObject/
      SegmentationClassAug/ # https://github.com/kazuto1011/deeplab-pytorch/blob/master/data/datasets/voc12/README.md
      # below are generated
      images_detectron2/
      annotations_detectron2/
        val/

It starts with a tar file VOCtrainval_11-May-2012.tar.

We use SBD augmentated training data as SegmentationClassAug following Deeplab

The directories images_detectron2 and annotations_detectron2 are generated by running

python datasets/prepare_voc_sem_seg.py

Expected dataset structure for D3:

$DETECTRON2_DATASETS/
  D3/
    d3_images/
    d3_json/
    d3_pkl/