{
  "class_mapping": {
    "0": {
      "category": 0,
      "name": "pedestrian"
    },
    "2": {
      "category": 1,
      "name": "car"
    },
    "5": {
      "category": 2,
      "name": "bus"
    },
    "7": {
      "category": 3,
      "name": "truck"
    }
  },
  "class_mapping_evidence": "Source class totals 0:27711, 2:103361, 5:4033, 7:3725 exactly match manuscript pedestrian/car/bus/truck totals. The manuscript does not print a numeric class-ID lookup table.",
  "mapping_evidence": "Paper explicitly lists A/B/C sequence splits; mapping to source directory names is reconstructed from 6/9/6 sequence counts and exact 89148/15820/33862 split box totals. Swapping the two six-sequence directories yields 100185/25266/13379 and fails.",
  "paper_sequence_splits": {
    "test": [
      "B4",
      "B6",
      "C2",
      "C3",
      "C4"
    ],
    "train": [
      "A1",
      "A2",
      "A3",
      "A4",
      "A5",
      "B1",
      "B2",
      "B5",
      "B8",
      "B9",
      "C1",
      "C5",
      "C6"
    ],
    "validation": [
      "A6",
      "B3",
      "B7"
    ]
  },
  "paper_title": "Understanding Dynamic Scenes at Gigapixel Scale: Wide-Area Spatio-Temporal Perception from UAVs",
  "scene_to_paper_prefix": {
    "NJU-xianlin": "C",
    "UAV-BASE-13": "A",
    "UAV-BASE-14": "B"
  },
  "source": "Locally supplied manuscript, page 10 Table 4; paper audit retained separately.",
  "tracking_evaluation": "Supplied manuscript protocol evaluates the five test sequences B4, B6, C2, C3, C4; this source release retains all 21 sequences."
}