benchmark/protocol.json

{
  "name": "Syn4D Multi-Task Benchmark",
  "version": "1.0",
  "input": {
    "root_env": "SYN4D_DATA_ROOT",
    "cluster_default_root": "/work/kelvin/Syn4D/subsets/kaggle_eval",
    "video": "<variant>/<scene>/mp4/<sequence>.mp4",
    "frames": [
      0,
      6,
      12,
      18,
      24,
      30,
      36,
      42,
      48,
      54,
      60,
      66,
      72,
      78,
      84,
      90,
      96,
      102,
      108,
      114,
      120,
      126,
      132,
      138,
      144,
      150,
      156,
      162,
      168,
      174,
      180,
      186
    ],
    "variants": [
      "og",
      "sim",
      "mixed",
      "mixed_no_bedlam"
    ],
    "scenes": [
      "antiquity",
      "dream",
      "gothic",
      "office",
      "outpost",
      "space",
      "street",
      "triplex"
    ],
    "cameras": [
      0,
      1
    ]
  },
  "tracking": {
    "ground_truth": "data/tracking_gt/<variant>/<scene>/<sequence>.npy",
    "ground_truth_schema": "syn4d-tracking-gt-v1 (structured NPY, allow_pickle=False)",
    "queries_per_sequence": 512,
    "query_frame": 0,
    "coordinate_system": "OpenCV frame-0 camera coordinates",
    "alignment": "one median scale per sequence",
    "thresholds_m": [
      0.1,
      0.3,
      0.5,
      1
    ],
    "dynamic_threshold_m": 0.01,
    "score": "0.5 * APD(all valid) + 0.5 * APD(dynamic valid)"
  },
  "depth": {
    "units": "meters",
    "alignment": "one median scale per sequence",
    "valid_range_m": [
      0.001,
      300
    ],
    "primary": "AbsRel",
    "reported": [
      "AbsRel",
      "SqRel",
      "RMSE",
      "RMSE-log",
      "SILog",
      "delta1",
      "delta2",
      "delta3"
    ]
  },
  "pose": {
    "representation": "camera-to-world 4x4 matrices",
    "alignment": "normalize to frame 0, then one Sim(3) per sequence",
    "primary": [
      "ATE RMSE",
      "RPE translation RMSE",
      "RPE rotation mean degrees"
    ],
    "reported": [
      "ATE",
      "position error",
      "rotation error",
      "RPE",
      "relative-pose AUC"
    ]
  },
  "aggregation": "macro average over sequences; also report by variant and scene"
}