🖼️ COOCO: Common Objects Out-of-Context

Authors: Filippo Merlo*, Ece Takmaz, Wenkai Chen, Albert Gatt *Corresponding author: [email protected]

Affiliations: Utrecht University, University of Trento

Status: Submitted to TACL (under review)

Code & Data: https://github.com/cs-nlp-uu/scenereg

---

📦 1. Dataset Overview

COOCO is a large-scale dataset designed to investigate how Vision-Language Models (VLMs) leverage scene context during referring expression generation. It focuses on semantic violations, evaluating model performance when target objects have low, medium, or high semantic relatedness to their surrounding scenes.

Each sample in COOCO contains:

• A scene from COCO-Search18 with a labeled target object, called Original version
• A version of the scene where the target object is just removed, called Clean version
• Multiple manipulated versions with inpainted objects of varying semantic relatedness

Each original image is augmented by replacing the target object with semantically related or unrelated alternatives. Relatedness is computed using ConceptNet embeddings and THINGSplus norms:

• Low
• Medium
• High
• Same-target: original object category generated (control)

Inpainting is guided by LLaVA-generated prompts and verified via an automated visual QA procedure.

---

The dataset supports the study of:

• Referring expression robustness to semantic incongruence

---

📁 2. Dataset Structure

Each sample in the COOCO dataset is a dictionary with the following fields:

• image (Image): The visual scene in RGB format.

• image_id (string): Unique identifier matching the image filename that is built following: {coco_id}_{scene}_{target}_{swapped_object}_relscore_{rel_level}.jpg

• scene (string): The high-level scene label (e.g., "cubicle office", "bakery shop"), drawn from a curated set of 29 categories.

• target (string): The original COCO object category that was removed from the image (e.g., "laptop").

• swapped_object (string): The object inserted in place of the target, selected based on semantic relatedness (e.g., "teapot").

• target_bbox (list[float32]): The bounding box coordinates of the original target object, in [x1, y1, w, h] format, normalized to the image size.

• rel_level (string): Semantic relatedness level between the scene and the swapped object. One of:

  • "low"
  • "medium"
  • "high"
  • "same_target"

• rel_score (float32): Cosine similarity score between the scene and swapped object label, computed from ConceptNet embeddings.

• excluded (bool): Indicates whether the sample was excluded during final quality verification (e.g., failed inpainting or object not correctly inserted).

Original and Clean image versions are distinguishable by the last part of their unique image_id.

---

🎯 3. Tasks & Experimental Use

COOCO supports:

• Object naming in context (referring expression generation)
• Robustness testing under:
  • Scene-object incongruity

---

🔍 4. Scene & Object Information

• Scenes: 29 curated categories from SUN397 (e.g., cubicle office, gas station)
• Objects: Selected from THINGSplus based on typicality and size match
• Spatial encoding: Normalized bounding box coordinates

---

📝 License

CC-BY-4.0

---

📬 Citation

@misc{merlo2025cooco,
  title={Common Objects Out-of-Context: Semantic Violation in Scenes for Investigating Multimodal Context in Referential Communication},
  author={Filippo Merlo and Ece Takmaz and Wenkai Chen and Albert Gatt},
  year={2025},
  url={https://arxiv.org/abs/2506.22274}
}

---