AHAAM/GUTIC

GUIC: Grounded and Ungrounded Image Confounders

Abstract

We introduce GUIC (Grounded and Ungrounded Image Confounders), a visual reasoning benchmark designed to evaluate the robustness of vision-language models under controlled in-image textual perturbations.

Modern multimodal models exhibit strong sensitivity to textual signals embedded within images (e.g., posters, apparel, overlays), often over-relying on such signals even when they are misleading, irrelevant, or ungrounded. GUIC systematically disentangles these effects by constructing structured image variants that isolate different classes of textual confounders.

---

1. Dataset Design

Each GUIC example consists of a question-image pair augmented with four structured variants:

1. Correct
2. Misleading Groundable
3. Misleading Ungroundable
4. Irrelevant

These variants differ only in the overlaid textual content, enabling controlled evaluation of textual influence.

1.1 Variant Definitions

• Correct
  The overlaid text matches the correct visual answer.

• Misleading Groundable
  The text refers to a visually present object that is not the correct answer.

• Misleading Ungroundable
  The text refers to an object not present in the scene.

• Irrelevant
  The text is unrelated to the scene and the question.

This structured design allows direct measurement of robustness across confounder types.

---

2. Annotations

GUIC provides two distinct spatial annotation types:

2.1 Object Bounding Box

Available for:

• correct_answer
• misleading_groundable

Stored as:

x, y, w, h

These coordinates localize the relevant visual object associated with the answer or grounded distractor.

---

2.2 Text Bounding Box

Available for all four variants.

Stored as:

[x1, y1, x2, y2]

This bounding box localizes the region containing the overlaid textual perturbation.

---

3. Dataset Structure

Each example contains:

• question_id
• image_id
• caption
• question
• seg_id
• notext (base image without overlay)
• seg_image (segmentation reference)

And four structured variant entries, each containing:

• text
• image
• bbox
• x, y, w, h (when applicable)

---

4. Splits

GUIC is released as a test-only benchmark.

DatasetDict({ test: 474 examples })

---

5. Limitations

• Text overlays are synthetic and controlled.
• Performance may depend on OCR capabilities.
• The benchmark isolates textual confounders but does not cover all multimodal adversarial phenomena.

---

5. Loading the Dataset

from datasets import load_dataset

ds = load_dataset("AHAAM/GUIC", split="test")

import random

idx = random.randint(0, 474)
example = ds[idx]

variants = [
    "correct_answer",
    "misleading_groundable",
    "misleading_ungroundable",
    "irrelevant_word"
]

def draw_boxes(image, text_bbox=None, object_bbox=None, object_color=None):
    img = image.copy()
    draw = ImageDraw.Draw(img)

    # 🔵 Text overlay box
    if text_bbox is not None:
        draw.rectangle(text_bbox, outline="blue", width=4)

    # Object box (green or red)
    if object_bbox is not None:
        x, y, w, h = object_bbox
        draw.rectangle([x, y, x + w, y + h], outline=object_color, width=4)

    return img


fig, axes = plt.subplots(1, 4, figsize=(20, 5))

for ax, variant in zip(axes, variants):
    img = example[variant]["image"]
    text_bbox = example[variant]["bbox"]
    text = example[variant]["text"]

    object_bbox = None
    object_color = None

    # 🟢 Correct object
    if variant == "correct_answer":
        object_bbox = (
            example[variant]["x"],
            example[variant]["y"],
            example[variant]["w"],
            example[variant]["h"]
        )
        object_color = "green"

    # 🔴 Incorrect grounded object
    elif variant == "misleading_groundable":
        object_bbox = (
            example[variant]["x"],
            example[variant]["y"],
            example[variant]["w"],
            example[variant]["h"]
        )
        object_color = "red"

    img_with_boxes = draw_boxes(
        img,
        text_bbox=text_bbox,
        object_bbox=object_bbox,
        object_color=object_color
    )

    ax.imshow(img_with_boxes)
    ax.set_title(f"{variant}\nText: {text}")
    ax.axis("off")

plt.tight_layout()
plt.show()