SRefcoco
收藏资源简介:
SRefCOCO是一个大规模多模态视觉定位数据集。它旨在突破传统视觉定位任务中仅依赖文本-图像双模态交互的局限,通过引入新颖的语音-文本-图像三模态数据,为具身AI模型提供端到端的听、读、看感知能力。该数据集基于经典的视觉定位数据集(如RefCOCO、RefCOCO+、RefCOCOg)构建并扩展。数据包含三个核心模态:1)图像:来自COCO数据集的RGB图像;2)文本:自然语言指代表达式;3)语音:通过Edge-TTS生成的16kHz采样率的.wav格式音频文件,这些音频经过了大量增强处理,以模拟复杂多变真实物理环境中的听觉场景。该数据集适用于需要融合听觉、语言理解和视觉感知的多模态交互与视觉定位研究。
SRefCOCO is a large-scale multimodal visual grounding dataset. It aims to break through the limitations of traditional visual grounding tasks that rely solely on text-image bimodal interactions by introducing novel speech-text-image trimodal data, providing embodied AI models with end-to-end hearing, reading, and seeing perceptual capabilities. The dataset is constructed and extended based on classic visual grounding datasets (such as RefCOCO, RefCOCO+, RefCOCOg). The data includes three core modalities: 1) Images: RGB images from the COCO dataset; 2) Text: natural language referring expressions; 3) Speech: .wav format audio files generated by Edge-TTS with a 16kHz sampling rate, which have undergone extensive enhancement processing to simulate auditory scenes in complex and varied real-world physical environments. This dataset is suitable for multimodal interaction and visual grounding research that requires the integration of auditory, language understanding, and visual perception.
SRefCOCO 数据集概述
SRefCOCO 是一个大规模多模态视觉定位数据集,其核心创新在于引入 语音-文本-图像 三元组,旨在赋予具身AI模型端到端的“听、读、看”感知能力。
- 数据来源:基于经典的视觉定位数据集(如 RefCOCO、RefCOCO+、RefCOCOg)构建。
- 数据组成:
- 图像:来自 COCO 数据集的标准 RGB 图像。
- 文本:自然语言指代表达式。
- 语音(音频):通过 Edge-TTS 生成、经过强增强处理的 16kHz
.wav音频文件,用于模拟复杂的真实物理环境。
- 许可协议:cc-by-nc-4.0。
- 使用方式:可通过 Hugging Face
datasets库加载,例如: python from datasets import load_dataset dataset = load_dataset("xutao2025/SRefcoco")




