raw_baseline_results.json

{
  "methodology_note": "Raw single-pass softmax baseline (dropout OFF, model.eval() fully -- NOT MC-Dropout), computed on the SAME held-out validation split (11666 frames / 1049 videos) as research_results.json's MC-Dropout numbers, so the two are directly comparable per architecture. Exists specifically to test H1 as literally worded in the proposal: 'MC Dropout will produce measurably better-calibrated confidence estimates (lower ECE) than each detector's raw softmax confidence.'",
  "val_frame_count": 11666,
  "val_video_count": 1049,
  "models": {
    "xception": {
      "frame_ece": 0.03287438389485061,
      "video_ece": 0.037745457105664085,
      "frame_auroc": 0.9976220768807692,
      "video_auroc": 0.9995995550611791
    },
    "sbi": {
      "frame_ece": 0.115351293327205,
      "video_ece": 0.13012384283197287,
      "frame_auroc": 0.9894134047671114,
      "video_auroc": 0.9981015943641083
    },
    "vit": {
      "frame_ece": 0.07730546174039567,
      "video_ece": 0.09054525756808249,
      "frame_auroc": 0.9829789807102295,
      "video_auroc": 0.9924063774564331
    },
    "lsda": {
      "frame_ece": 0.06318750835388509,
      "video_ece": 0.07276670377292158,
      "frame_auroc": 0.9929955925425256,
      "video_auroc": 0.9984723767148683
    }
  }
}