US 12,094,124 B2
Method for measuring the boundary performance of a semantic segmentation network
Ze Guo, Stuttgart (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Oct. 2, 2020, as Appl. No. 17/062,092.
Claims priority of application No. 19201614 (EP), filed on Oct. 7, 2019.
Prior Publication US 2021/0101287 A1, Apr. 8, 2021
Int. Cl. G06T 7/12 (2017.01); B25J 9/16 (2006.01); G06F 18/21 (2023.01); G06F 18/2431 (2023.01); G06N 3/08 (2023.01); G06V 10/28 (2022.01); G06V 10/34 (2022.01); G06V 10/44 (2022.01); G06V 10/776 (2022.01); G06V 20/10 (2022.01); G06V 20/56 (2022.01)
CPC G06T 7/12 (2017.01) [B25J 9/163 (2013.01); B25J 9/1697 (2013.01); G06F 18/2163 (2023.01); G06F 18/217 (2023.01); G06F 18/2431 (2023.01); G06N 3/08 (2013.01); G06V 10/28 (2022.01); G06V 10/34 (2022.01); G06V 10/44 (2022.01); G06V 10/776 (2022.01); G06V 20/10 (2022.01); G06V 20/56 (2022.01)] 21 Claims
OG exemplary drawing
 
1. A method for measuring a boundary performance of a semantic segmentation network, the method comprising:
determining an original boundary mask defining a boundary between different classes of an image, wherein determining the original boundary mask includes thresholding an image map having a foreground and a background;
determining a thickened boundary mask dependent on the original boundary mask using morphological operations;
determining a true positive rate, a false positive rate and a false negative rate, dependent on the original boundary mask and/or the thickened boundary mask;
determining a final boundary metric, indicating the boundary performance, dependent on the determined true positive rate, the determined false positive rate, and the determined false negative rate; and
determining a boundary loss function as a weighted sum of a first loss term and a second loss term, wherein the first loss term represents a crossentropy loss for semantic segmentation, and wherein the second loss term includes a differentiable boundary loss term;
wherein the differentiable boundary loss term=αLce+(1−α)Lb, and wherein α is a hyper parameter weighting factor, wherein Lce is the first loss term representing a crossentropy loss for semantic segmentation, and wherein Lb is the second loss term that is the differentiable boundary loss term,
wherein the morphological operations include erosion, which shrinks an area boundary, and dilation, which enlarges the area boundary, wherein the erosion and the dilation are performed using structuring element filters, and
wherein the true positive rate relates to a number of pixels correctly classified as foreground, wherein the false positive rate relates to a number of pixels falsely classified as foreground, and wherein the false negative rate relates to a number of pixels falsely classified as background.