Dense mapping from sparse visual odometry: a lightweight uncertainty-guaranteed depth completion method

Introduction: Visual odometry (VO) has been widely deployed on mobile robots for spatial perception. State-of-the-art VO offers robust localization, the maps it generates are often too sparse for downstream tasks due to insufffcient depth data. While depth completion methods can estimate dense depth from sparse data, the extreme sparsity and highly uneven distribution of depth signals in VO (∼ 0.15% of the pixels in the depth image available) poses signiffcant challenges. Methods: To address this issue, we propose a lightweight Image-Guided Uncertainty-Aware Depth Completion Network (IU-DC) for completing sparse depth from VO. This network integrates color and spatial information into a normalized convolutional neural network to tackle the sparsity issue and simultaneously outputs dense depth and associated uncertainty. The estimated depth is uncertainty-aware, allowing for the filtering of outliers and ensuring precise spatial perception. Results: The superior performance of IU-DC compared to SOTA is validated across multiple open-source datasets in terms of depth and uncertainty estimation accuracy. In real-world mapping tasks, by integrating IU-DC with the mapping module, we achieve 50 × more reconstructed volumes and 78% coverage of the ground truth with twice the accuracy compared to SOTA, despite having only 0.6 M parameters (just 3% of the size of the SOTA). Discussion: Our code will be released at https://github.com/YangDL-BEIHANG/Dense-mapping-from-sparse-visual-odometry/tree/d5a11b4403b5ac2e9e0c3644b14b9711c2748bf9.