VistaQA: Benchmarking Joint Visual Question Answering and Pixel-Level Evidence

Overview

VistaQA is a benchmark for the joint evaluation of free-form answer correctness and pixel-level visual evidence alignment in visual question answering. It contains 1,157 expert-curated samples across six task types and six visual domains, spanning perception to compositional and relational reasoning. Each sample requires both a textual answer and corresponding segmentation masks that support the prediction. The benchmark also includes hallucination-aware samples in which no valid visual evidence exists.

---

Dataset Structure

VistaQA/
    ├── 1.jpg
    ├── 1.json
    ├── 2.jpg
    ├── 2.json
    └── ...

Each image file (.jpg or .png) is paired with a corresponding .json file sharing the same file ID.

---

Annotation Format (Example)

{
  "image": {
    "image_id": 979,
    "width": 1500,
    "height": 2060,
    "file_name": "979.jpg"
  },
  "question": "how many windows on the building are not partially occluded by the balusters?",
  "answer": "there are 13 windows that not partially occluded by the balusters.",
  "task_type": "counting",
  "task_domain": "outdoor",
  "num_instances": 13,
  "hallucination": 0,
  "annotations": [
    {
      "id": 523353741,
      "segmentation": {
        "size": [
          2060,
          1500
        ],
        "counts": "l][T1n0g4TOPe1e6I6M3N10000O2O0000000000000O100001O00000000000000000000000000000000000000000000000001O0000000000O1000000001O0000O2O00000000000001O0O1000000000O101O0001O00O1000O1001N10000O1O100O1O1O1O1O1N2O1O1O1_N^VNXLbi1f3cVNVL^i1i3gVNRLZi1m3SWNfKoh1X4TWNfKlh15ZVNV3k0dLmh1OdVNV3`0kLmh1JoVNR35SMQi1BZWNo2F^Mjj14VTN]1Q1^Nmj1OZTNHKV1P1SOoj1IjTNl0;ZO\\l1:hSNE`n1O2Lcejb1"
      },
      "bbox": [
        578.0,
        636.0,
        112.0,
        228.0
      ],
      "area": 23791
    }
  ]
}
   

Note: For brevity, only one of the 13 masks is shown.

Field Descriptions

• image: Filename of the associated image and its metadata (e.g., width, height)
• question: Visual question answering (VQA)
• answer: Ground-truth answer (free-form)
• task_type: Type of reasoning (attribute, identification, OCR, counting, spatial, comparison)
• task_domain: Domain category (AV, indoor, outdoor, robotics, math, science)
• num_instances: Number of instances for visual evidence masks
• hallucination: Indicates whether valid visual evidence exists (0 = evidence present, 1 = no valid evidence)
• annotations: Segmentation mask(s) representing supporting evidence