{
  "name": "CAP",
  "path": "/home/aviad/datasets/CAP",
  "texture_type": "category-independent appearance partitioning (two-region transition crops)",
  "k_nominal": "2",
  "source": "RWTD_Coco CAP_balanced_256: 212 manually curated anchors plus 44 CAP-supervised additions, all from the COCO-Stuff 164K stuffthingmaps tiers; NOT derived from ADE20K",
  "built_utc": "2026-09-02T08:45:51.767195+00:00",
  "selection": {
    "method": "manual review",
    "selection_dir": "(local manual selection)",
    "n_selected": 256,
    "by_source_pool": {
      "CAP_500": 152,
      "CAP_500/candidate": 13,
      "CAP_FC/candidate": 8,
      "raw/tier1": 49,
      "raw/tier2": 34
    },
    "by_tier": {
      "tier1": 201,
      "tier2": 55
    },
    "n_cap500": 248,
    "n_cap_cf": 8,
    "by_curation": {
      "CAP_supervised_extension": 44,
      "manual anchor": 212
    },
    "anchor_note": "The manually curated anchors are identified per entry by provenance.curation.is_cap_anchor=true; the rest were added by the CAP-supervised ranker. No separate anchor directory is kept -- filter on that flag instead.",
    "note": "Entries with cap_selection.cap_cf=true come from the CAP-CF pool and were chosen because their two regions look alike. They are counterfactual probes, not typical samples; keep them separable when computing set-level statistics."
  },
  "directories": {
    "images": 256,
    "masks": 256,
    "overlays": 256,
    "textures_mask": 512
  },
  "metadata_files": {
    "metadata.json": {
      "num_entries": 256,
      "total_texture_instances": 512,
      "non_empty_descriptions": 0
    }
  },
  "k_distribution": {
    "2": 256
  },
  "description_policy": "blank (annotation_status=pending_human)",
  "provenance_audit": {
    "ade20k_derived_entries": 0,
    "clean": true,
    "image_bytes_verified_against_source": 256
  },
  "mirrors": {
    "repo_copy": "RWTD_Coco/output/CAP_balanced_256",
    "note": "same 256 samples and the same region masks. This copy follows the Detecture dataset layout: masks/<id>_annotation.png is the RGB annotation the sibling datasets use. The repo copy follows the RWTD_Coco tier layout instead -- the same regions live there as label_masks/<id>.png (tri-value) and masks/<id>.png (binary, region A) -- and it adds manifest.jsonl/.csv, source_label_crops/ and contact_sheets/ for the statistics pipeline."
  },
  "image_sizes": {
    "sampled": 50,
    "unique_sizes_in_sample": 4,
    "width": {
      "min": 192,
      "max": 256
    },
    "height": {
      "min": 192,
      "max": 256
    },
    "constant": false,
    "example_size": [
      192,
      192
    ]
  }
}