BAR (Biased Action Recognition)

有偏见的动作识别（BAR）数据集是一个真实世界的图像数据集，分为六个动作类别，这些类别偏向于不同的地方。作者通过检查 imSitu 来解决这六个动作类别，imSitu 提供来自 Google 图像搜索的静态动作图像以及动作和地点标签。详细地说，作者选择了动作类别，其中每个候选动作的图像具有共同的地点特征。同时，动作类候选的位置特征应该是不同的，以便仅从位置属性对动作进行分类。选择的配对是六个典型的动作地点配对：（攀岩，岩壁），（潜水，水下），（钓鱼，水面），（赛车，APavedTrack），（投掷，运动场）和（跳马，天空）。The Biased Action Recognition (BAR) dataset is a real-world image dataset composed of six action categories that are biased towards distinct locations. To curate these six action categories, the authors utilized the imSitu dataset, which provides static action images sourced from Google Image Search along with action and location labels. Specifically, the authors selected action categories where images of each candidate action share common location features, while the location features of different action candidates should be distinct, enabling action classification to be performed solely based on location attributes. The selected pairs are six typical action-location pairs: (rock climbing, rock face), (diving, underwater), (fishing, water surface), (racing, APavedTrack), (throwing, sports field), and (vaulting, sky).

搜集汇总

背景与挑战

背景概述

BAR数据集是一个包含六个动作类别的图像数据集，每个动作类别与特定的地点特征相关联，旨在研究动作识别中的偏见问题。该数据集由韩国科学技术院于2020年发布，相关研究发表在NeurIPS 2020上。