具方向性学习的宽基线相对相机位姿估计本文提出 DirectionNet 模型，通过预测 5D 相对位姿空间上的离散分布来改进相机位姿回归，实现将相机姿态分解为 3D 方向向量并在球面上估算离散分布。我们从 Matterport3D 和 InteriorNet 构建的具有挑战性的综合和真实姿态估计数据集中评估了我们的模型，结果表明我们的方法比直接回归大大减少了误差。(Translation: This paper proposes DirectionNet model to improve camera pose regression by predicting a discrete distribution over the 5D relative pose space. DirectionNet factorizes relative camera pose to a set of 3D direction vectors and estimates the distribution on the sphere, resulting in a significant error reduction compared to direct regression methods, as evaluated on challenging synthetic and real pose estimation datasets constructed from Matterport3D and InteriorNet.)
Jun, 2021