Anywhere3D-Bench
收藏资源简介:
Anywhere3D-Bench是一个全面的3D视觉定位基准数据集,包含2632个引用表达式和3D边界框对,涵盖了四个不同的定位级别:人类活动区域、超出对象的未占用空间、场景中的单个对象以及细粒度的对象部分。该数据集由BIGAI、清华大学、北京大学和北京理工大学的研究人员创建,旨在评估和推动3D视觉定位模型在多级别场景下的定位能力,特别是对于超出对象级别的空间区域和细粒度对象部分的定位。数据集来源于ScanNet、MultiScan、3RScan和ARKitScenes的验证集,通过人类编写的提示和GPT-4生成引用表达式,并由人工进行标注和验证,确保每个引用表达式都能精确地定位到一个目标3D边界框。
Anywhere3D-Bench is a comprehensive 3D visual grounding benchmark dataset containing 2632 pairs of referring expressions and 3D bounding boxes, covering four distinct grounding levels: human activity areas, unoccupied spaces beyond objects, individual objects in the scene, and fine-grained object parts. Developed by researchers from BIGAI, Tsinghua University, Peking University, and Beijing Institute of Technology, this dataset aims to evaluate and advance the grounding capabilities of 3D visual grounding models across multi-level scenarios, with a particular focus on spatial regions beyond object-level boundaries and fine-grained object parts. The dataset is derived from the validation splits of ScanNet, MultiScan, 3RScan, and ARKitScenes. Referring expressions were generated using human-written prompts and GPT-4, followed by manual annotation and verification to ensure that each referring expression can accurately pinpoint a corresponding target 3D bounding box.
- 1From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes清华大学, 北京大学, 北京理工大学 · 2025年



