Computer Vision News - July 2021

Joseph Lambourne 37 Best of CVPR 2021 to use to describe the initial geometry there. They can then be passed through a multi-level perceptron (MLP), which is like a learning set of weights, to start to recognize patterns in these local structures. Then the learned representation is attached to this representation of the same coedge but in the next layer of the network. This can be done for every coedge in the model and then max pooling used to generate new embeddings for the faces, which can be passed on to build a multi-layer architecture. This is the core idea of BRepNet . The team have another accepted paper at CVPR this year, called UV-Net: Learning From Boundary Representations . This proposes a way to provide more geometric information about B -rep models to neural networks by describing the actual surface geometry using regular grids of points which are regularly spaced in the UV parametrization of the surfaces. Currently, BRepNet uses very simple input features to describe the geometry and uses the Fusion 360 Gallery Segmentation Dataset t o help test the power of these algorithms using real-world rather than synthetic data. Moving forward, they want to combine the point grid concept in the UV-Net architecture , which is a good way of describing the geometry of the faces, with the BRepNet convolution scheme, which is an efficient way of doing convolutions on the boundary representation. This will provide a better way to understand the models and help users of AutoCAD software to automate many of the tedious selections that they are making now. “ Obviously with BRepNet we have a very flexible way of defining which entities are used in each convolution, ” Joe tells us. “ When we did the experiment, the convolution kernel which gave the best performance turned out to include exactly the entities which were present in the winged edge data structure invented by Baumgart back in 1972. This shows how classic geometric modelling ideas can play a part in modern machine learning! ” This area of technology is growing in popularity, with new publications showing the possibility to build B -rep models using sequential data, and transformer models. “ The DeepCAD paper published recently on arXiv has everyone very excited because they see the potential to not only automate things like manufacturing, but to be able to create new designs as B-rep models as well, ” Joe says. “ It’s a really exciting area for people wanting to explore something new in machine learning to get involved with. ” If you are interested in getting started working with solid models, you will find the public code for BRepNet here. To learn more about Joe’s work, you can view his presentation and PDF poster anytime, and take part in his live Q&A during Paper Session Ten today (Thursday) at 11:00-13:30 EDT. "This shows how class c geometric modeling ideas can play a part in modern machine learning!" o o o to us to describe the initi l geometry ere. Th y ca then be passed through a multi-level perceptron (MLP), which is like a learning set of w ights, to star to recognize patterns in these local structures. Then learned representation is attached to this repres nt tion of he same coedg but in the next layer of the network. This can be done for every coedge in the model and then max pooling used to generate new embeddings for the faces, which can be passed on to build a multi-layer architecture. This is the core idea of BRepNet . The t am have another accepted paper at CVPR this year, called UV-Net: L arning From Boundary Repres nt ti s . This proposes a way to provide more geometric information about B -rep models to neural networks by describing the ac tual surface geo etry using regular g rids of points which are regularly spaced in t he UV parametrization of the surfaces. Currently, BRepNet uses very simple input features to des ribe geom ry and uses the Fu ion 360 G llery S gme ta ion Dataset t help test the pow r of th se algorithms using real-world rather than synthetic ata. M ving f rward, y wa t to combine the point grid concept in the UV-Net rchitectur , which is a good way of describing the geometry of the faces, with the BRepNet convolution scheme, which is an efficient way of doing convolutions on the boundary automate many of the tedious selections that they ar making ow. “ Obviously with BRepNet we have a very flexible way of defining which entities are used in each convolution, ” Joe tells us. “ When we did the experim nt, he convolution k rnel which gave the best performance turned out to include exactly the entities which were present in the winged edge data structure invented by Bau gart back in 1972. This shows how classic geometric modelling ideas can play a part in modern machine learning! ” This area of technology is growing in popularity, with new publications showing the possibility to build B -rep models using sequential data, and transformer models. “ The DeepCAD paper published recently on arXiv has everyone very excited because they see the potential to not only automate things like manufacturing, but to be able to create new d signs as B-rep models as ell, ” Joe says. “ It’s a really exciting area for people wanting to explore something new in machine learning to get involved with. ” If you are interested in getting started working with solid models, you will find the public code for BRepNet here. To learn more about Joe’s work, you can view his presentation and PDF poster anytime, and take part in his