Evaluation of deep learning algorithms for semantic segmentation of car parts

Abstract Evaluation of car damages from an accident is one the most important processes in insurance business. Currently, it still needs a manual examination every basic part. It expected that smart device will be able to do this evaluation more efficiently future. In study, we evaluated and compared five deep learning algorithms for semantic segmentation parts. The baseline reference algorithm was Mask R-CNN, other were HTC, CBNet, PANet, GCNet. Runs instance conducted with those algorithms. HTC ResNet-50 best on various kinds cars such as sedans, trucks, SUVs. achieved mean average precision at 55.2 our original data set, assigned different labels left right sides 59.1 when single label both sides. addition, models tested robustness, by running them images parts, real environment weather conditions, including snow, frost, fog lighting conditions. GCNet robust; performance under corruption, mPC = 35.2, relative degradation corrupted data, clean (rPC), 64.4%, labels, 38.1 rPC $$69.6\%$$ 69.6 % left- right-side parts considered same findings study may directly benefit developers automated damage system their quest design.