RefSpatial-Bench
收藏资源简介:
RefSpatial-Bench 是一个基于真实世界杂乱场景的基准测试,旨在评估更复杂的具有推理能力的多步空间指示。它包括三个任务:位置、放置和未见过。位置任务要求模型预测一个二维点,指示唯一的靶标对象。放置任务要求模型在所需空闲空间内预测一个二维点。未见过集合由来自位置/放置任务的77个样本组成,专门设计用于评估模型在 RefSpatial 上的 SFT/RFT 训练后的泛化能力,因为它包括 RefSpatial 中不存在的新的空间关系组合。数据集还引入了推理步骤(`step`),它是帮助约束搜索空间的锚点对象及其空间关系的数量。较高的 `step` 值反映了更大的推理复杂性和对空间理解和推理的更强需求。数据集提供两种格式:Hugging Face 数据集格式和原始数据格式。Hugging Face 数据集格式位于 `data/` 文件夹中,包含 HF 兼容的分片:`location`、`placement` 和 `unseen`。每个样本包括以下字段:`id`、`object`、`prompt`、`suffix`、`image`、`mask` 和 `step`。原始数据格式在 `Location/`、`Placement/` 和 `Unseen/` 文件夹中提供,每个文件夹包含 RGB 图像、地面实况二进制掩码和包含指示性提示和元数据的 `question.json` 文件。每个条目在 `question.json` 文件中都有以下格式:`id`、`object`、`prompt`、`suffix`、`rgb_path`、`mask_path`、`category` 和 `step`。
RefSpatial-Bench is a benchmark built on real-world cluttered scenes, aiming to evaluate more complex multi-step spatial instructions with reasoning capabilities. It encompasses three tasks: Location, Placement, and Unseen. The Location task requires the model to predict a 2D point that indicates the unique target object. The Placement task requires the model to predict a 2D point within the required free space. The Unseen split consists of 77 samples sourced from the Location and Placement tasks, specifically designed to assess the generalization performance of models after SFT/RFT training on RefSpatial, as it contains novel spatial relation combinations that do not exist in the rest of RefSpatial. The dataset also introduces the `step` metric, which refers to the number of anchor objects and their spatial relations that assist in constraining the search space. A higher `step` value indicates greater reasoning complexity and a stronger requirement for spatial understanding and reasoning. The dataset provides two formats: the Hugging Face Dataset format and the raw data format. The Hugging Face Dataset format is stored in the `data/` folder, including HF-compatible splits: `location`, `placement`, and `unseen`. Each sample contains the following fields: `id`, `object`, `prompt`, `suffix`, `image`, `mask`, and `step`. The raw data format is offered in the `Location/`, `Placement/`, and `Unseen/` folders, with each folder housing RGB images, ground-truth binary masks, and a `question.json` file that contains instructional prompts and metadata. Each entry in the `question.json` file follows the format: `id`, `object`, `prompt`, `suffix`, `rgb_path`, `mask_path`, `category`, and `step`.




