Frozen detector · independent annotations · visible failures

What does the detector catch?
What does it miss?

A small held-out probe of real video: 12 DAVIS sequences, 36 frame pairs, and a detector frozen before the first prediction. Reference labels come from human-drawn foreground masks.

34 / 36

positive foreground changes detected

2 missed

inspect every miss below

0 negatives

real-video false-positive rate cannot be estimated from this sample

Detection recall: 94.4% on this proxy. This is not overall accuracy, semantic incident detection, or an official DAVIS score. All 36 selected pairs happen to be positive. Constructed controls have their own denominator below.

Inspect all 36 real probes

Misses appear first. Expand a case to inspect the independent annotation and exact detector evidence.

36 cases

Missed drift chicane frames 0 → 5 1.23% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for drift chicane
Reference
Foreground occupancy changed by 1.23% of image area.
Detector
0 card(s). OpenCV found no frame transition above the configured threshold.
Next action
complete_no_material_change
Inspect exact evidence
[]
Missed drift chicane frames 20 → 25 0.56% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for drift chicane
Reference
Foreground occupancy changed by 0.56% of image area.
Detector
0 card(s). OpenCV found no frame transition above the configured threshold.
Next action
complete_no_material_change
Inspect exact evidence
[]
Detected bike packing frames 0 → 5 3.60% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bike packing
Reference
Foreground occupancy changed by 3.60% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.024172,
    "changed_regions": [
      [
        276,
        86,
        348,
        394
      ],
      [
        259,
        200,
        50,
        280
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "20879fef6c2430d13851707baa4371bd8b854e0d65225571ceb1f675d19c135a",
    "previous_sha256": "13731d5dcf2f1e6e70278a06fba7ecc9f7110ed7834686efc69a8b69d26be1b7",
    "timestamp_ms": 1000
  }
]
Detected bike packing frames 10 → 15 9.65% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bike packing
Reference
Foreground occupancy changed by 9.65% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.046191,
    "changed_regions": [
      [
        241,
        0,
        397,
        480
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "035f59781c9d6799302e6d7256e5c575e351009e3c452c35bfdcb6511a0e6bfa",
    "previous_sha256": "d4b4645128431caaffe9d4a8462f23b2670c6437aa18ed28d091a1a52995d3c1",
    "timestamp_ms": 1000
  }
]
Detected bike packing frames 20 → 25 9.09% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bike packing
Reference
Foreground occupancy changed by 9.09% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.052557,
    "changed_regions": [
      [
        223,
        38,
        657,
        442
      ],
      [
        304,
        453,
        104,
        27
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "1dd93aa3d3d6e16cbaca5b133b6fb8a17e1ebf2563cf27cfb8e4033390b083ad",
    "previous_sha256": "9a2ba088b0f1c183eb1daf9c18282455199269d653606b258b6c3b3451731a59",
    "timestamp_ms": 1000
  }
]
Detected blackswan frames 0 → 5 3.91% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for blackswan
Reference
Foreground occupancy changed by 3.91% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.078438,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        704,
        139,
        150,
        61
      ],
      [
        243,
        168,
        97,
        35
      ],
      [
        544,
        268,
        232,
        176
      ],
      [
        719,
        384,
        65,
        33
      ],
      [
        786,
        410,
        68,
        70
      ],
      [
        674,
        427,
        62,
        53
      ],
      [
        726,
        435,
        48,
        45
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "b88f2ec6e695d40b193d0ef038457df20f47dc14dc74f30403c90cf7883fd1e0",
    "previous_sha256": "6544f8b98ca68fdd88f6ce5c706a8a6ad71e07c574eb87f0eacabd04549ff6f8",
    "timestamp_ms": 1000
  }
]
Detected blackswan frames 10 → 15 3.95% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for blackswan
Reference
Foreground occupancy changed by 3.95% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.092804,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        62,
        152,
        110,
        46
      ],
      [
        564,
        277,
        90,
        65
      ],
      [
        651,
        315,
        80,
        51
      ],
      [
        571,
        323,
        82,
        97
      ],
      [
        528,
        361,
        48,
        78
      ],
      [
        662,
        375,
        62,
        66
      ],
      [
        392,
        445,
        62,
        35
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "fe21230a575922e86609a95ca2d7c1b750f6684ccdad90e76b8b0dc74e361b07",
    "previous_sha256": "3dd8fce0d295abd39f889cc36a009515be7a6d323c5b779a672c5be82a03646a",
    "timestamp_ms": 1000
  }
]
Detected blackswan frames 20 → 25 2.60% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for blackswan
Reference
Foreground occupancy changed by 2.60% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.081924,
    "changed_regions": [
      [
        0,
        0,
        854,
        417
      ],
      [
        795,
        167,
        59,
        31
      ],
      [
        588,
        216,
        76,
        63
      ],
      [
        0,
        334,
        283,
        107
      ],
      [
        546,
        351,
        86,
        57
      ],
      [
        718,
        367,
        38,
        56
      ],
      [
        0,
        411,
        84,
        50
      ],
      [
        548,
        422,
        84,
        58
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "41f538b0c0c3631ab71f392097329116a633054f4f5accd8cdda4506c7aa8ae8",
    "previous_sha256": "925f484d908b4c069b75151eead42e7988356a4abf7affcadf950bfc7f532319",
    "timestamp_ms": 1000
  }
]
Detected bmx trees frames 0 → 5 2.74% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bmx trees
Reference
Foreground occupancy changed by 2.74% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.122571,
    "changed_regions": [
      [
        0,
        0,
        278,
        49
      ],
      [
        0,
        0,
        854,
        480
      ],
      [
        0,
        263,
        79,
        56
      ],
      [
        235,
        376,
        165,
        104
      ],
      [
        0,
        444,
        51,
        36
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5c729c95e5aec3de03b075d5d6b3e681e1b730061d5d252dff7e557cae9848f6",
    "previous_sha256": "85c9cfbb96c6d29ba931dc5f856aa802248e6908abb25072f8a3791a2a3d419f",
    "timestamp_ms": 1000
  }
]
Detected bmx trees frames 10 → 15 2.82% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bmx trees
Reference
Foreground occupancy changed by 2.82% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.12493,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        239,
        0,
        117,
        37
      ],
      [
        0,
        226,
        210,
        109
      ],
      [
        127,
        255,
        95,
        38
      ],
      [
        0,
        295,
        68,
        40
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "a65511d84b8145211ab4671f24c36d234d6c5d2d28719bb737ad81e197b3cd01",
    "previous_sha256": "029a8678aa8be969bbea1bb77ea1d1158553045664f97afc74ff8d3bee383f33",
    "timestamp_ms": 1000
  }
]
Detected bmx trees frames 20 → 25 2.91% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for bmx trees
Reference
Foreground occupancy changed by 2.91% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.14026,
    "changed_regions": [
      [
        0,
        0,
        39,
        77
      ],
      [
        0,
        0,
        206,
        326
      ],
      [
        0,
        0,
        854,
        480
      ],
      [
        632,
        0,
        222,
        192
      ],
      [
        120,
        64,
        83,
        127
      ],
      [
        716,
        456,
        138,
        24
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5755d34fc66970709c98111a0e32c8d1590f7c736dbd9d313b3fdc95c8590ced",
    "previous_sha256": "660904c9861ab739f43d98529839e302ec85507f51d0861e7c087a955cb40cc9",
    "timestamp_ms": 1000
  }
]
Detected breakdance frames 0 → 5 6.33% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for breakdance
Reference
Foreground occupancy changed by 6.33% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.05001,
    "changed_regions": [
      [
        97,
        0,
        324,
        459
      ],
      [
        0,
        112,
        82,
        177
      ],
      [
        784,
        146,
        66,
        59
      ],
      [
        433,
        196,
        55,
        90
      ],
      [
        0,
        285,
        119,
        163
      ],
      [
        445,
        306,
        53,
        111
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "8dceede9fbf006de0a1cda112185925eb0c8a2f67c990e759abda0630030d5c3",
    "previous_sha256": "f3b630cadeb8e8ab011a00126d7cbcdaafe30bfd5371d711752fcef534c2d99e",
    "timestamp_ms": 1000
  }
]
Detected breakdance frames 10 → 15 9.16% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for breakdance
Reference
Foreground occupancy changed by 9.16% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.06353,
    "changed_regions": [
      [
        0,
        91,
        573,
        389
      ],
      [
        0,
        117,
        120,
        117
      ],
      [
        719,
        125,
        58,
        90
      ],
      [
        661,
        182,
        59,
        101
      ],
      [
        388,
        192,
        100,
        68
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "e4bdc733c10e39f60589c5484f96c7e1d17b54438b5797fec495f4215dc94bbb",
    "previous_sha256": "67eae5e23b921da85aca74ff34827731c02a343082661da896fbb6d03f76ed2e",
    "timestamp_ms": 1000
  }
]
Detected breakdance frames 20 → 25 9.61% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for breakdance
Reference
Foreground occupancy changed by 9.61% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.055779,
    "changed_regions": [
      [
        0,
        32,
        601,
        423
      ],
      [
        719,
        103,
        67,
        112
      ],
      [
        0,
        122,
        93,
        138
      ],
      [
        648,
        169,
        78,
        76
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "8e8afff708355d656345aadcac9d1375a19dd5772b61f855beb21d014f0c8fe1",
    "previous_sha256": "54b8022e9a377089087134f7297a5da685f38ae081a26e9dcd9f8b544b6a7c60",
    "timestamp_ms": 1000
  }
]
Detected camel frames 0 → 5 7.27% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for camel
Reference
Foreground occupancy changed by 7.27% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.105559,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        0,
        372,
        76,
        46
      ],
      [
        771,
        394,
        83,
        41
      ],
      [
        767,
        446,
        87,
        24
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "9a9d54427a94e4f0b4fb8f414906027bc53eb1914f107617ad7ae58e3e9545ce",
    "previous_sha256": "10d0a7dc4e35df3c069c49b74e5a9c6cc8862f67ff0a230aeec683d3084dda95",
    "timestamp_ms": 1000
  }
]
Detected camel frames 10 → 15 4.45% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for camel
Reference
Foreground occupancy changed by 4.45% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.08573,
    "changed_regions": [
      [
        0,
        0,
        36,
        72
      ],
      [
        0,
        0,
        854,
        480
      ],
      [
        710,
        412,
        144,
        33
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5d4df96ff0eac412b9ce264e0e0346fe80b15b8af160d6b20f8dded36dec4509",
    "previous_sha256": "dbaba96f67c70b93e4d2a0228f938e63a2d059f09abad7fd9ca14ca95ac5448a",
    "timestamp_ms": 1000
  }
]
Detected camel frames 20 → 25 6.38% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for camel
Reference
Foreground occupancy changed by 6.38% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.113612,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        669,
        400,
        133,
        27
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "3ce8d3abe271b61deecb79f80635c9e86f2948432071220bf708a2058238608b",
    "previous_sha256": "e72d368d9bb8b1ff5fcae9f04dafc0ef172a0e198c41cb33e65cef4b3e4749df",
    "timestamp_ms": 1000
  }
]
Detected car roundabout frames 0 → 5 2.43% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car roundabout
Reference
Foreground occupancy changed by 2.43% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.13338,
    "changed_regions": [
      [
        0,
        0,
        854,
        358
      ],
      [
        490,
        282,
        113,
        70
      ],
      [
        0,
        359,
        376,
        42
      ],
      [
        587,
        408,
        267,
        15
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "4a13b3093eb0f5cb7240c7b8b4818435db9c022a8e7133ab8697470d979cb742",
    "previous_sha256": "62e3dbf66819620208a9207a85cb87f519bd03149d0360e952796ac9ea0ac9e8",
    "timestamp_ms": 1000
  }
]
Detected car roundabout frames 10 → 15 0.78% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car roundabout
Reference
Foreground occupancy changed by 0.78% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.122312,
    "changed_regions": [
      [
        0,
        0,
        854,
        359
      ],
      [
        0,
        389,
        683,
        28
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5dacbffc6e01ed16232b40671762a941bf5c2f8b842559cbb9c235df4b7b90d6",
    "previous_sha256": "2b8f460883b34e16b3dcb253d7e620627b20057f992bc4f1210bf2f3bec31ac4",
    "timestamp_ms": 1000
  }
]
Detected car roundabout frames 20 → 25 0.68% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car roundabout
Reference
Foreground occupancy changed by 0.68% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.129362,
    "changed_regions": [
      [
        0,
        0,
        854,
        352
      ],
      [
        314,
        273,
        326,
        88
      ],
      [
        148,
        288,
        369,
        78
      ],
      [
        0,
        409,
        584,
        14
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "d66d00b8364de66ae604aac2faaee47ae64a73c3df7ddda6510993f3ec91655d",
    "previous_sha256": "d9edda987f014b271498c872c834e60a4b6f5af7f89455f5e42fa6d897abdeb8",
    "timestamp_ms": 1000
  }
]
Detected car shadow frames 0 → 5 4.69% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car shadow
Reference
Foreground occupancy changed by 4.69% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.088511,
    "changed_regions": [
      [
        0,
        0,
        135,
        36
      ],
      [
        0,
        0,
        854,
        333
      ],
      [
        164,
        0,
        115,
        53
      ],
      [
        280,
        0,
        59,
        61
      ],
      [
        114,
        69,
        31,
        70
      ],
      [
        143,
        69,
        141,
        65
      ],
      [
        0,
        357,
        335,
        19
      ],
      [
        518,
        362,
        261,
        15
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5210b807cb709228f34ca6b03e58eea45912c48f92ce87adb525de4941b6f5c4",
    "previous_sha256": "d03fa65f5b71caffc525d35357c537b4e18ec785204691eb61e57de3cdd43126",
    "timestamp_ms": 1000
  }
]
Detected car shadow frames 10 → 15 1.55% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car shadow
Reference
Foreground occupancy changed by 1.55% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.086843,
    "changed_regions": [
      [
        0,
        0,
        854,
        375
      ],
      [
        119,
        0,
        196,
        52
      ],
      [
        348,
        0,
        49,
        62
      ],
      [
        411,
        4,
        75,
        72
      ],
      [
        272,
        77,
        63,
        63
      ],
      [
        716,
        317,
        138,
        45
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "a95952946cbfce8495b4c53859b7bc5995753f6b75549a046d136696706f80d2",
    "previous_sha256": "5b3c592c0908ae6c78b4e08223830cbb8d572490604e718191dd0c7df583bda2",
    "timestamp_ms": 1000
  }
]
Detected car shadow frames 20 → 25 1.16% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for car shadow
Reference
Foreground occupancy changed by 1.16% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.09699,
    "changed_regions": [
      [
        0,
        0,
        854,
        378
      ],
      [
        171,
        0,
        86,
        45
      ],
      [
        262,
        0,
        181,
        59
      ],
      [
        475,
        0,
        49,
        68
      ],
      [
        539,
        13,
        37,
        62
      ],
      [
        576,
        19,
        41,
        64
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "456458914f9cb8b1a7658ba79b5aff3e829386b3cbca4f04a7300a6fae3647ce",
    "previous_sha256": "879e91783bb123b9b45d5b087032312ecd2e76413ab10a92041903d72de9d062",
    "timestamp_ms": 1000
  }
]
Detected cows frames 0 → 5 3.16% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for cows
Reference
Foreground occupancy changed by 3.16% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.071568,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        431,
        0,
        89,
        58
      ],
      [
        626,
        325,
        228,
        155
      ],
      [
        596,
        438,
        47,
        42
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "894ecfad43293c30f033f4653c9fd228be8934c8dbcd531571747fe94cec2182",
    "previous_sha256": "720a86212910a3359e277f56fe7dfa4ce7e772dfbbb0ac3b0237f33c210eb7b9",
    "timestamp_ms": 1000
  }
]
Detected cows frames 10 → 15 1.71% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for cows
Reference
Foreground occupancy changed by 1.71% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.040738,
    "changed_regions": [
      [
        0,
        25,
        854,
        455
      ],
      [
        784,
        331,
        70,
        43
      ],
      [
        675,
        394,
        148,
        86
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "656a4ff568c86b225b41f0ae19d63fa3f7ad8269fe4b6fd7bfa75ea6d6a35a5d",
    "previous_sha256": "576f0b1c8b9358f8b0920c7e483c4223213c4ac33f9ed23e56c81b73428ea826",
    "timestamp_ms": 1000
  }
]
Detected cows frames 20 → 25 4.15% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for cows
Reference
Foreground occupancy changed by 4.15% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.084235,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "5f23ccbd77ce8803d67829961070b86bb1a8a90f3bc9580bf37ee2e5bff938e8",
    "previous_sha256": "e74bb2c752662f96f6210d3cd47dddc21d2085af818a6e8c615f4130f83dd7fe",
    "timestamp_ms": 1000
  }
]
Detected dance twirl frames 0 → 5 4.06% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dance twirl
Reference
Foreground occupancy changed by 4.06% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.10948,
    "changed_regions": [
      [
        0,
        0,
        854,
        437
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "bc7622af6e484890835b45136cafaaea2fea73fd1ada7d6c879629e6cf2e2f53",
    "previous_sha256": "36685274794859438def6a0cddee0ff47dc8fc69c2b3310a60e1b42661abf0bb",
    "timestamp_ms": 1000
  }
]
Detected dance twirl frames 10 → 15 5.56% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dance twirl
Reference
Foreground occupancy changed by 5.56% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.130528,
    "changed_regions": [
      [
        0,
        0,
        854,
        473
      ],
      [
        381,
        386,
        53,
        43
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "a4fc3631083271651aab2816b1ac5cad84d90fbc97f09cde5ce69ba582470500",
    "previous_sha256": "a99f62fed3c836261233a5e43b43fedbb211100cd16de6bc3c6ba0f4c88d593b",
    "timestamp_ms": 1000
  }
]
Detected dance twirl frames 20 → 25 4.85% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dance twirl
Reference
Foreground occupancy changed by 4.85% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.121997,
    "changed_regions": [
      [
        0,
        0,
        854,
        431
      ],
      [
        363,
        404,
        68,
        55
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "6b096645b528fab3555f45e39f96af245b017e2c362d1dab59adbe0086c549e8",
    "previous_sha256": "c97544aee98a8c86bd5b63f6662b17166e8c1ab920f492c3c6575eb29639fcbf",
    "timestamp_ms": 1000
  }
]
Detected dog frames 0 → 5 7.58% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dog
Reference
Foreground occupancy changed by 7.58% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.087323,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        221,
        0,
        133,
        60
      ],
      [
        356,
        0,
        133,
        72
      ],
      [
        0,
        51,
        149,
        74
      ],
      [
        762,
        68,
        92,
        86
      ],
      [
        124,
        233,
        82,
        57
      ],
      [
        569,
        433,
        97,
        47
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "6645c6ebeac82ba30a08452c7bfabfe9c819f327ff1469b6837d6b806a0aabf7",
    "previous_sha256": "883d5e7c1f808e8604e787cc7177fbae91c6ffaff19a8a1256f21a2e1595fe62",
    "timestamp_ms": 1000
  }
]
Detected dog frames 10 → 15 6.46% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dog
Reference
Foreground occupancy changed by 6.46% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.084508,
    "changed_regions": [
      [
        0,
        0,
        854,
        480
      ],
      [
        689,
        62,
        134,
        105
      ],
      [
        0,
        77,
        129,
        68
      ],
      [
        0,
        241,
        190,
        227
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "6850325721eaa160ff52bd5ee9af992492eaff8032145b9eb94ab26671ecbeb3",
    "previous_sha256": "6cf7b71dc6b732d57e376284f90f5aed9141862e671497a416758d4755ac0fbe",
    "timestamp_ms": 1000
  }
]
Detected dog frames 20 → 25 7.23% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dog
Reference
Foreground occupancy changed by 7.23% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.098204,
    "changed_regions": [
      [
        0,
        0,
        299,
        30
      ],
      [
        0,
        0,
        854,
        480
      ],
      [
        756,
        42,
        95,
        107
      ],
      [
        0,
        250,
        305,
        228
      ],
      [
        131,
        262,
        63,
        59
      ],
      [
        306,
        370,
        80,
        76
      ],
      [
        387,
        375,
        66,
        47
      ],
      [
        804,
        416,
        50,
        64
      ],
      [
        162,
        425,
        71,
        55
      ],
      [
        41,
        429,
        55,
        51
      ],
      [
        316,
        444,
        110,
        36
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "2a84b45a7764cc6fa2394792df0473b8622f5b60a0f18ba642ca8fda21425d4a",
    "previous_sha256": "fd0677f57a9da6a1e84872bafa2a5006466728c7d158bf160d7870927f71a357",
    "timestamp_ms": 1000
  }
]
Detected dogs jump frames 0 → 5 2.54% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dogs jump
Reference
Foreground occupancy changed by 2.54% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.029792,
    "changed_regions": [
      [
        0,
        126,
        854,
        331
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "f05395bc46f3627fe18449d656bf37283f9419e7c93c06db6ceba1f92e6c706e",
    "previous_sha256": "2bd517ae8a17ddfe98b28d471978065cb1dec3f551d924fd6caebad3254f8489",
    "timestamp_ms": 1000
  }
]
Detected dogs jump frames 10 → 15 9.03% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dogs jump
Reference
Foreground occupancy changed by 9.03% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.053739,
    "changed_regions": [
      [
        0,
        139,
        854,
        248
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "a31d3d316c04d4f8baea820e6a45cfc273331465bc31fe847efebb3e5aff0430",
    "previous_sha256": "687c86236f1562ee5339b5f78618c53caf38948f29af38153a8455e1170cdb12",
    "timestamp_ms": 1000
  }
]
Detected dogs jump frames 20 → 25 5.56% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for dogs jump
Reference
Foreground occupancy changed by 5.56% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.031946,
    "changed_regions": [
      [
        0,
        150,
        541,
        237
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "0c4b3db68327b61023962afc2fcb0535ccdf1b175ea2191d3bb48b0342e1729b",
    "previous_sha256": "3b00775954315169592708d14e1a423672d94495b1046d31ed23c638503f84dd",
    "timestamp_ms": 1000
  }
]
Detected drift chicane frames 10 → 15 1.29% foreground change

Human annotation: before (left), after (right). Orange boxes show the detector's regions projected onto the annotation. These diagrams contain no original camera pixels.

DAVIS foreground annotation pair for drift chicane
Reference
Foreground occupancy changed by 1.29% of image area.
Detector
1 card(s). OpenCV found localized changes that are reviewable without an immediate escalation.
Next action
seal_evidence_for_agent_summary
Inspect exact evidence
[
  {
    "change_score": 0.024109,
    "changed_regions": [
      [
        570,
        25,
        172,
        168
      ],
      [
        277,
        199,
        186,
        63
      ]
    ],
    "frame_index": 1,
    "frame_sha256": "8493f7109c77a7474774b62c8c2a6773b19d7c57e4e31eb0bdd7736e2a08b1d0",
    "previous_sha256": "77cfe2b78d10cffd4664bee96d5cbfa3d3ac1a628fed60825b092b93cc8fd84c",
    "timestamp_ms": 1000
  }
]

How nuisance changes behave

36 constructed controls based on real frames. Foreground geometry stays fixed. Lighting changes are still pixel changes, so a trigger is a nuisance under this proxy, not a broken pixel-difference calculation.

Exact repeat

The same real frame twice.

0 / 12

nuisance triggers under the foreground proxy

Small sensor-like noise

Seeded ±2 pixel intensity perturbation; constructed control.

0 / 12

nuisance triggers under the foreground proxy

Lighting change

+40 intensity with clipping; geometry unchanged. Constructed control.

12 / 12

nuisance triggers under the foreground proxy

Operational limit: changing illumination can trigger review even when foreground geometry has not moved. Small foreground movement can also be missed. Use these cards for assisted review; do not treat a quiet result as proof that nothing happened.

How to interpret the evidence

Sampling, labels, and leakage controls

First 12 alphabetically sorted names from DAVIS 2017 validation; fixed frame pairs (0,5), (10,15), (20,25). No selection by score. Human foreground occupancy XOR ≥0.5% is positive; ≤0.1% is negative; the middle interval is unscored. There were 0 ambiguous pairs.

Detector commit d09d96889f30d07e0902830b90cb616a60d4149e; event threshold 0.02, reset threshold 0.01. Pair probes reset hysteresis each time. Sequence-macro recall is 94.4% across 12 sequences; pairs within a sequence are correlated. This public validation subset was unused in this project's prior tuning but is not a private test set. No thresholds were adjusted after seeing results.

Reproduce and verify the receipt
python fetch_evaluation_data.py --output-dir /tmp/davis-probes
python evaluate_heldout.py --data-dir /tmp/davis-probes --output-dir /tmp/heldout-results
python build_evaluation_page.py --results /tmp/heldout-results/results.json

Source files are checksum-verified. All per-pair predictions, source hashes, dependency versions, and the protocol hash are in results.json. The receipt hashes canonical JSON excluding its own receipt_sha256 field.

Receipt: 64ee986bf1bcd1f5949b7064d6d5754c05b1aa73592f10e5a7b80aac3fef3c0e

What this evaluation does not establish

No object identity, intent, incident classification, production accuracy, localization IoU, real-world review-time savings, or AWS latency is measured here. The annotations define a different target from global pixel intensity. Camera and background motion can disagree with that target. Future tuning on these examples requires a new holdout.