SQA3D: Situated Question Answering in 3D Scenes (ICLR 2023, https://arxiv.org/abs/2210.07474)

1. Download the SQA3D dataset under assets/data/. The following files should be used:

./assets/data/sqa_task/balanced/*
./assets/data/sqa_task/answer_dict.json

2. The dataset has been splited into train, val and test. For each category, we offer both question file, ex. v1_balanced_questions_train_scannetv2.json, and annotations, ex. v1_balanced_sqa_annotations_train_scannetv2.json

• The format of question file: Run the following code:

  import json
  q = json.load(open('v1_balanced_questions_train_scannetv2.json', 'r'))
  # Print the total number of questions
  print('#questions: ', len(q['questions']))
  print(q['questions'][0])
  

  The output is:

  {
    "alternative_situation": 
      [
        "I stand looking out of the window in thought and a radiator is right in front of me.",
        "I am looking outside through the window behind the desk."
      ],
    "question": "What color is the desk to my right?",
    "question_id": 220602000000,
    "scene_id": "scene0380_00",
    "situation": "I am facing a window and there is a desk on my right and a chair behind me."
  }
  

  The following fileds are useful: question, question_id, scene_id, situation.

• The format of annotations: Run the following code:

  import json
  a = json.load(open('v1_balanced_sqa_annotations_train_scannetv2.json', 'r'))
  # Print the total number of annotations, should be the same as questions
  print('#annotations: ', len(a['annotations']))
  print(a['annotations'][0])
  

  The output is

  {
    "answer_type": "other",
    "answers": 
      [
        {
          "answer": "brown", 
          "answer_confidence": "yes", 
          "answer_id": 1
        }
      ],
    "position": 
      {
        "x": -0.9651003385573296,
        "y": -1.2417634435553606,
        "z": 0
      },
    "question_id": 220602000000,
    "question_type": "N/A",
    "rotation": 
      {
        "_w": 0.9950041652780182,
        "_x": 0,
        "_y": 0,
        "_z": 0.09983341664682724
      },
    "scene_id": "scene0380_00"
  }
  

  The following fields are useful: answers[0]['answer'], question_id, scene_id. Note: To find the answer of a question in the question file, you need to use lookup with question_id.

3. We provide the mapping between answers and class labels in answer_dict.json

import json
j = json.load(open('answer_dict.json', 'r'))
print('Total classes: ', len(j[0]))
print('The class label of answer \'table\' is: ', j[0]['table'])
print('The corresponding answer of class 123 is: ', j[1]['123'])

4. Loader, model and training code can be found at https://github.com/SilongYong/SQA3D