Datasets

Prepare the local COCO and Visual Genome files used by Mirage.

Mirage reads dataset files already available on disk. The demo does not download full datasets automatically. By default, paths are resolved relative to the repository root; use --annotations and --images to point to another location.

COCO

The default validation layout is:

datasets/coco/
├── annotations/
│   └── instances_val2017.json
└── val2017/
    ├── 000000000001.jpg
    └── ...

The annotation file must use COCO instance format with images, annotations, and categories entries. Choose another split with --split; the default is val2017.

Visual Genome

The default layout is:

datasets/visual_genome/
├── objects.json
└── VG_100K/
    ├── <image files>
    └── ...

Override paths

Use explicit paths when files are stored outside the default directories:

python demo.py \
  --dataset coco \
  --metric chair \
  --annotations /data/coco/annotations/instances_val2017.json \
  --images /data/coco/val2017 \
  --samples 10

Keep large datasets, model weights, and generated results outside version control. See the repository’s .gitignore for the ignored local data and cache paths.