Ego-HOIBench
收藏资源简介:
Ego-HOIBench是一个新的数据集,旨在促进Ego-HOI检测的基准和开发。该数据集包含超过27K个高质量的 egocentric 图像,具有123个细粒度的手-动词-对象三元组注释,覆盖了日常生活活动中的丰富场景、对象类型和手部配置。此外,该数据集还定义了两种Ego-HOIBench挑战,以探索Ego-HOI检测任务。为了建立一个新的基线,我们提出了一个轻量级且有效的交互增强方案,即HGIR,该方案利用手部姿态和几何线索来从全局角度改善交互表示。我们的方法可以灵活地与现成的HOI检测器集成,无需额外的手部姿态估计器,即可实现出色的效率。实验结果表明,我们的方法在Ego-HOIBench上取得了显著的性能提升。
Ego-HOIBench is a novel dataset designed to facilitate benchmarking and development of Ego-HOI detection. This dataset contains over 27,000 high-quality egocentric images, with annotations covering 123 fine-grained hand-verb-object triplet categories, and encompasses diverse scenarios, object types and hand configurations across daily life activities. Additionally, this dataset defines two Ego-HOIBench-specific challenges to advance the exploration of the Ego-HOI detection task. To establish a new baseline, we propose a lightweight yet effective interaction enhancement scheme named HGIR, which leverages hand pose and geometric cues to enhance interaction representations from a holistic perspective. Our method can be flexibly integrated with off-the-shelf HOI detectors without requiring additional hand pose estimators, delivering excellent computational efficiency. Experimental results demonstrate that our method achieves significant performance improvements on the Ego-HOIBench dataset.

- 1Egocentric Human-Object Interaction Detection: A New Benchmark and Method香港理工大学电子与电气工程系 · 2025年



