[Docs] Meaning of "tokens_positive" in the annotations

[OasisYang]

Branch

main branch https://mmdetection3d.readthedocs.io/en/latest/

📚 The doc issue

Hi, thanks for releasing this great dataset.

I'm confused about the meaning of "tokens_positive" which refers to "Position indices of mentioned objects in the text."

• What are the position indices? Are they object IDs?
• For several examples, I cannot find any object mentioned in the text. For the following example, could you point out the mentioned objects?
  {
  "sub_class": "VG_Single_Attribute_Unique",
  "scan_id": "matterport3d/ZMojNkEp431/region0",
  "target_id": [
  48
  ],
  "distractor_ids": [],
  "text": "The item with a black metallic sheen belongs to hardware and has a minimalistic design.",
  "target": [
  "object"
  ],
  "anchors": [],
  "anchor_ids": [],
  "tokens_positive": {
  "48": [
  [
  0,
  36
  ]
  ]
  },
  "ID": "VG_Single_Attribute_Unique__matterport3d/ZMojNkEp431/region0__2"
  },

Thanks for your help!

Suggest a potential alternative/fix

No response

[rbler1234]

Thank you for your feedback! The explanation for this part in the README.md is not very detailed. tokens_positive represents the indices of the key phrase in the sentence that denotes the target object, which is usually the subject of the sentence along with any accompanying adjectives. For example, in the sentence "The item with a black metallic sheen belongs to hardware and has a minimalistic design," the segment from the 0th to the 36th position, "The item with a black metallic sheen," is the key phrase representing the target object.
As part of the ground truth, tokens_positive is invisible to the model during inference and is only used to assist in training processes.