ABSTRACT Direct visual odometry (DVO) is an important vision task which aims to obtain the camera motion via minimizing the photometric error across the different correlated images. However, the previous research on DVO rarely considered the motion bias and only calculated using single direction, therefore potentially ignoring useful information compared with leveraging diverse directions. We assume that jointly considering forward and backward calculation can improve the accuracy of pose estimation. To verify our assumption and solid this contribution, in this paper, we test various combination of direct dense methods, including different error metrics, e.g., (intensity, gradient magnitude), alignment strategies (Forward-Compositional, Inverse-Compositional), and calculation directions (forward, backward, and bi-direction). We further study the issue of motion bias in RGB-D visual odometry and propose four strategy options to improve pose estimation accuracy, e.g., joint bi-direction estimation; two stage bi-direction estimation; transform average with weights; and transform fusion with covariance. We demonstrate the effectiveness and efficiency of our proposed algorithms across a range of popular datasets, e.g., TUM RGB-D and ICL-NUIM, in which we achieve an impressive performance through comparing with state of the art methods and provide benefits for existing RGB-D visual odometry and visual SLAM systems.

中文翻译：

---

双向直接RGB-D视觉里程计

摘要 直接视觉里程计 (DVO) 是一项重要的视觉任务，旨在通过最小化不同相关图像之间的光度误差来获得相机运动。然而，之前对 DVO 的研究很少考虑运动偏差，只使用单一方向计算，因此与利用不同方向相比，可能会忽略有用的信息。我们假设联合考虑前向和后向计算可以提高姿态估计的准确性。为了验证我们的假设并巩固这一贡献，在本文中，我们测试了直接密集方法的各种组合，包括不同的误差度量，例如（强度、梯度幅度）、对齐策略（正向组合、逆组合）和计算方向（向前、向后和双向）。我们进一步研究了 RGB-D 视觉里程计中的运动偏差问题，并提出了四种提高姿态估计精度的策略选项，例如联合双向估计；两阶段双向估计；用权重变换平均值；并用协方差变换融合。我们在一系列流行数据集（例如 TUM RGB-D 和 ICL-NUIM）中证明了我们提出的算法的有效性和效率，其中我们通过与最先进的方法进行比较获得了令人印象深刻的性能，并为现有的 RGB- D 视觉里程计和视觉 SLAM 系统。