{
  "name": "POVBench",
  "version": "1.0.0",
  "description": "A benchmark for contextual observer grounding: localizing a target object from an observer-relative natural-language sentence, under three levels of observer specification.",
  "paper": {
    "title": "Contextual Observer Grounding: Evaluating Situated Spatial Reasoning in Vision-Language Models",
    "venue": "EMNLP 2026 Findings",
    "project_page": "https://mimo-owl.github.io/POVBench/",
    "code": "https://github.com/mimo-owl/povbench-eval"
  },
  "generated_at": "2026-09-06",
  "statistics": {
    "houses": 47,
    "scenarios": 288,
    "instances_per_condition": {
      "A_inferred": 813,
      "B_stated": 813,
      "C_given": 813
    },
    "instances_total": 2439,
    "avg_valid_directions_per_scenario": 2.82,
    "direction_coverage": {
      "front": {
        "count": 280,
        "pct_of_scenarios": 97.2
      },
      "up": {
        "count": 189,
        "pct_of_scenarios": 65.6
      },
      "left": {
        "count": 178,
        "pct_of_scenarios": 61.8
      },
      "right": {
        "count": 166,
        "pct_of_scenarios": 57.6
      }
    },
    "observer_anchor_categories_realized": 25,
    "observer_anchor_categories_candidate": 31,
    "target_categories_realized": 25,
    "images": {
      "exploration_images": 1278,
      "img_from_observer": 419,
      "gt_images": 588,
      "top_down_view": 47,
      "total": 2332
    },
    "image_resolution": "512x512"
  },
  "conditions": {
    "A_inferred": "stage5[<direction>].type_a  - the observer's furniture is not named",
    "B_stated": "stage5[<direction>].type_b  - the observer's furniture is named",
    "C_given": "stage6[<direction>].type_c  + the observer-viewpoint image at rgb_path"
  },
  "ground_truth": {
    "field": "placements[<direction>].position",
    "space": "3D world coordinates (AI2-THOR)",
    "note": "2D image coordinates are not stored; they are obtained by pinhole projection using the per-scenario `camera` field. See the evaluation code."
  },
  "provenance": {
    "scene_generator": "ProcTHOR",
    "simulator": "AI2-THOR",
    "renderer": "CloudRendering",
    "judge_vlm": "Gemma (used only to judge directional plausibility and support surface)",
    "build_seed": 42,
    "note": "Houses were drawn from a pre-generated ProcTHOR pool; build_seed applies to the benchmark construction, not to house generation. The generation pipeline is not released, and the plausibility judgments come from a VLM, so the build is not bit-reproducible from this seed."
  },
  "license": {
    "annotations": "CC-BY-4.0",
    "rendered_images_and_scene_metadata": "Apache-2.0 (derived from AI2-THOR / ProcTHOR)"
  },
  "house_ids": [
    "train_house_00000",
    "train_house_00001",
    "train_house_00002",
    "train_house_00003",
    "train_house_00004",
    "train_house_00005",
    "train_house_00006",
    "train_house_00007",
    "train_house_00008",
    "train_house_00009",
    "train_house_00010",
    "train_house_00012",
    "train_house_00013",
    "train_house_00014",
    "train_house_00015",
    "train_house_00016",
    "train_house_00018",
    "train_house_00019",
    "train_house_00020",
    "train_house_00021",
    "train_house_00022",
    "train_house_00023",
    "train_house_00024",
    "train_house_00025",
    "train_house_00026",
    "train_house_00027",
    "train_house_00028",
    "train_house_00029",
    "train_house_00031",
    "train_house_00032",
    "train_house_00033",
    "train_house_00034",
    "train_house_00035",
    "train_house_00036",
    "train_house_00037",
    "train_house_00038",
    "train_house_00040",
    "train_house_00041",
    "train_house_00042",
    "train_house_00043",
    "train_house_00044",
    "train_house_00045",
    "train_house_00046",
    "train_house_00048",
    "train_house_00049",
    "train_house_00051",
    "train_house_00053"
  ],
  "per_house_scenario_counts": {
    "train_house_00000": 5,
    "train_house_00001": 2,
    "train_house_00002": 9,
    "train_house_00003": 9,
    "train_house_00004": 7,
    "train_house_00005": 8,
    "train_house_00006": 11,
    "train_house_00007": 5,
    "train_house_00008": 13,
    "train_house_00009": 6,
    "train_house_00010": 13,
    "train_house_00012": 5,
    "train_house_00013": 13,
    "train_house_00014": 4,
    "train_house_00015": 7,
    "train_house_00016": 5,
    "train_house_00018": 4,
    "train_house_00019": 8,
    "train_house_00020": 8,
    "train_house_00021": 3,
    "train_house_00022": 8,
    "train_house_00023": 1,
    "train_house_00024": 5,
    "train_house_00025": 13,
    "train_house_00026": 6,
    "train_house_00027": 8,
    "train_house_00028": 9,
    "train_house_00029": 6,
    "train_house_00031": 6,
    "train_house_00032": 6,
    "train_house_00033": 2,
    "train_house_00034": 1,
    "train_house_00035": 5,
    "train_house_00036": 2,
    "train_house_00037": 5,
    "train_house_00038": 5,
    "train_house_00040": 3,
    "train_house_00041": 1,
    "train_house_00042": 6,
    "train_house_00043": 8,
    "train_house_00044": 1,
    "train_house_00045": 3,
    "train_house_00046": 6,
    "train_house_00048": 6,
    "train_house_00049": 12,
    "train_house_00051": 8,
    "train_house_00053": 1
  },
  "note": "Regenerated from the released artifacts. The original build summaries recorded only one shard of a 3-way parallel build and are not included."
}