metadata
license: cc-by-nc-4.0
📢 SRefCOCO: A Large-Scale Multi-Modal Visual Grounding Dataset
📖 Introduction
Visual grounding aims to locate target regions in an image based on natural language instructions. However, mainstream paradigms are largely restricted to "text-image" bimodal interactions, severely limiting their application flexibility in highly dynamic real-world physical scenarios.
SRefCOCO breaks this traditional text-only constraint by introducing a novel Speech-Text-Image triplet dataset. It is designed to endow embodied AI models with end-to-end "listen, read, and look" perceptual capabilities. The dataset is built upon the classic visual grounding datasets (e.g., RefCOCO, RefCOCO+, RefCOCOg) and extends them with robust acoustic instructions.
- Images: Standard RGB images from the COCO dataset.
- Texts: Natural language referring expressions.
- Speech (Audio): 16kHz
.wavaudio files generated via Edge-TTS and heavily augmented to simulate complex real-world physical environments.
🚀 How to Use
You can easily load the dataset using the Hugging Face datasets library:
from datasets import load_dataset
# Load the entire dataset
dataset = load_dataset("xutao2025/SRefcoco")
# Print the first training sample
print(dataset['train'][0])