DATASET.md

Preparing Dataset

Our dataloader follows Detectron2 contains (1) A dataset registrator. (2) A dataset mapper. (3) A dataset loader. We modify the dataset registrator and mapper for different datasets.

Training Dataset Note

There are overlap between COCO2017, COCO-Karpathy and REF-COCO dataset, and ref-coco is all overalp with the COCO2017 training data, we have exclude the refcocog-umd validation, coco-karpathy test split during training.

ADE20K, Cityscapes, COCO

Please Refer to Mask2Former.

BDD100K

Please download the 10k split of BDD100k at https://doc.bdd100k.com/download.html#id1

Expected dataset structure for cityscapes:

.
└── bdd100k/
    ├── images/
    │   └── 10k/
    │       ├── test
    │       ├── train
    │       └── val
    └── labels/
        ├── ins_seg
        ├── pan_seg
        └── sem_seg

RefCOCO

Please download the original refcoco datasets at https://github.com/lichengunc/refer.

Expected dataset structure for refcoco:

.
└── refcocoseg/
    └── refcocog/
        ├── instances.json
        ├── refs(google).p
        └── refs(umd).p

Also download the coco dataset at https://cocodataset.org/#home:

Expected dataset structure for coco:

.
└── coco/
    ├── annotations
    ├── train2017
    └── val2017

After preparing the dataset, run the following command:

# NOTE: Please modify coco_root and ref_root
python3 refcoco2json.py

SUN-RGBD

SCAN-Net