DATASET.md

DATASET

We adopt the same protocol with BGNN for Visual Genome and Openimage datasets.

Visual Genome

The following is adapted from BGNN by following the same protocal of Unbiased Scene Graph Generation from Biased Training You can download the annotation directly by following steps.

Download:

1. Download the VG images part1 part2. Extract these images to the file /path/to/vg/VG_100K.

2. Download the scene graphs annotations and extract them to /path/to/vg/vg_motif_anno.

3. Link the image into the project folder

ln -s /path-to-vg datasets/vg

Openimage V4/V6

We adopt Openimage datasets from BGNN.

Download

1. The initial dataset(oidv6/v4-train/test/validation-annotations-vrd.csv) can be downloaded from offical website.

2. The Openimage is a very large dataset, however, most of images doesn't have relationship annotations. To this end, we filter those non-relationship annotations and obtain the subset of dataset (.ipynb for processing ).

3. You can download the processed dataset: Openimage V6(38GB), Openimage V4(28GB)

4. By unzip the downloaded datasets, the dataset dir contains the images and annotations folder. Link the open_imagev4 and open_image_v6 dir to the /datasets/openimages then you are ready to go.

mkdir datasets/openimages
ln -s /path/to/open_imagev6 datasets/openimages

Customized Dataset

Register you implement Dataset and Evaluator by editing the cvpods/data/datasets/paths_route.py.

_PREDEFINED_SPLITS_VG_STANFORD_SGDET = {
    "dataset_type": "VGStanfordDataset",  # visual genome stanford split
    "evaluator_type": {
        "vgs": "vg_sgg",
    },
    "vgs": {
      # the former is image directry path, the later is annotation directry path
        "vgs_train": ("vg/VG_100k_images", "vg/vg_motif_anno"),
        "vgs_val": ("vg/VG_100k_images", "vg/vg_motif_anno"),
        "vgs_test": ("vg/VG_100k_images", "vg/vg_motif_anno"),

    }
}

More details refer to cvpods tutorial.