Instance Segmentation Sample Clauses

Instance Segmentation. The first experiment evaluates the performance of our instance segmentation approach. Its outcome supports the claim that we can accurately delineate trees along with their canopies into individual instances. We show comparisons against the methods of Xxxxx et al. [36], Xxxxxxxx et al. [5], Xx et al. [18], and Xxxxxxx et al. [7]. The motivation for these baselines is that, in part, they are state-of-the- art methods, but also that their working implementation is available open-source. We do not compare against the results of Xxxxxxxxx et al. [16] since their method does not perform canopy separation. For the forest plots under study, DBH measurements were collected only for a subset of trees in each plot. We labeled only this subset of trees, resulting in a dataset containing partially annotated point clouds. To then evaluate the segmentation performance of the approaches, we use the standard panoptic quality (PQ) metric [4], [15]. PQ is given by Method Plot 1 Plot 2 Plot 3 Plot 4 Average Xxxxxxx et al. [7] 2.988 2.721 2.963 - 2.890 Xxxxxxxxx et al. [16] 0.031 0.051 0.100 0.021 0.051 Ours 0.088 0.116 0.182 0.054 0.110 TABLE II: Comparison of DBH estimation performance in RMSE (m) of different algorithms. Our method, however, segments also the tree’s canopy along with the stem into each instance. Moreover, since we do not have semantic information, we potentially segment the bushes and undergrowth vegetation along with the canopy. To mitigate this, we clip the height-normalized point cloud below 1 m as a pre-processing step to remove such vegetation but we still found it insufficient in the forest plots that we studied. Simply clipping more of the cloud also renders the DBH estimation inefficient, as this is typically measured at around 1.3 m. As a consequence, our cylinder fitting strategy is not restricted to the stem, which degrades the performance in terms of the DBH estimation. Xxxxxx, however, that the gap in estimation accuracy is, on average, below 6 cm. The other geometric method proposed by Xxxxxxx et al. [7] fails on this data with an average RMSE of 2.89 m. We exclude the result for Plot 4 from the average and in Tab. II since, in this case, Xxxxxxx et al. [7] estimate an average DBH of over 300 meters. This is due to many outliers in their tree instance segmentation. We notice that, in general, the rather low performance of Xxxxxxx et al. [7] stems from their Σ (p,g)∈TP IoU(p, g)