Abstract
Chamfer Distance (CD) and Earth Mover's Distance (EMD) are two broadlyadopted metrics for measuring the similarity between two point sets. However,CD is usually insensitive to mismatched local density, and EMD is usuallydominated by global distribution while overlooks the fidelity of detailedstructures. Besides, their unbounded value range induces a heavy influence fromthe outliers. These defects prevent them from providing a consistentevaluation. To tackle these problems, we propose a new similarity measure namedDensity-aware Chamfer Distance (DCD). It is derived from CD and benefits fromseveral desirable properties: 1) it can detect disparity of densitydistributions and is thus a more intensive measure of similarity compared toCD; 2) it is stricter with detailed structures and significantly morecomputationally efficient than EMD; 3) the bounded value range encourages amore stable and reasonable evaluation over the whole test set. We adopt DCD toevaluate the point cloud completion task, where experimental results show thatDCD pays attention to both the overall structure and local geometric detailsand provides a more reliable evaluation even when CD and EMD contradict eachother. We can also use DCD as the training loss, which outperforms the samemodel trained with CD loss on all three metrics. In addition, we propose anovel point discriminator module that estimates the priority for another guideddown-sampling step, and it achieves noticeable improvements under DCD togetherwith competitive results for both CD and EMD. We hope our work could pave theway for a more comprehensive and practical point cloud similarity evaluation.Our code will be available at:https://github.com/wutong16/Density_aware_Chamfer_Distance .