For structural component segmentation we propose and implement superpixel classification and unsupervised semantic segmentation based on feature clustering for detailed refinement.
This approach best exploits properties of the provided dataset, including the limited texture library used for synthetic data generation. The approach was very easy to implement with aid of open python libs and free license of ROBOFLOW.ai
Another advantage of our method is that no downscaling of images was done. (no downsize for training)
(Plz use TF 2.6.0 some weird model loading problems otherwise)