artifacts/vg_5k/split_half.json


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

{
  "protocol": "Each scene's parts split into disjoint halves within one modality; a relation field is built from each half and the two are matched by the anchor bound. Both halves describe the same scenes, so this upper-bounds any cross-modal result.",
  "results": {
    "text": {
      "scenes": 256,
      "mean_parts": 15.9765625,
      "split_half_correlation": 0.7933697484313135,
      "split_half_anchor_bound": 0.490625
    },
    "vision": {
      "scenes": 256,
      "mean_parts": 6.0,
      "split_half_correlation": 0.9836462733134468,
      "split_half_anchor_bound": 0.99375
    }
  },
  "cross_modal_bound_for_reference": 0.291
}