Techniques for analyzing vehicle design deviations using deep learning with neural networks
A design application is configured to generate a latent space representation of a fleet of pre-existing vehicles. The design application encodes vehicle designs associated with the fleet of pre-existing vehicles into the latent space representation to generate a first latent space location. The first latent space location represents the characteristic style associated with the fleet of pre-existing vehicles. The design application encodes a sample design provided by a user into the latent space representation to produce a second latent space location. The design application then determines a distance between the first latent space location and the second latent space location. Based on the distance, the design application generates a style metric that indicates the aesthetic similarity between the sample design and the vehicle designs associated with the fleet of pre-existing vehicles. The design application can also generate new vehicle designs based on the latent space representation and the sample design.
1. A computer-implemented method for generating vehicle designs, the method comprising:
encoding, using an encoder network, a first vehicle design into a first location within a latent space representation of vehicle design features;
encoding, using the encoder network, a second vehicle design into a second location within the latent space representation;
generating a first metric based on the first location and the second location, wherein the first metric indicates a degree to which the first vehicle design differs from a characteristic style associated with the second vehicle design;
traversing the latent space representation from the second location towards the first location to a third location in the latent space representation, wherein a distance traversed is determined based on the first metric; and
generating, using a generator network, a third vehicle design based on the third location by increasing a dimensionality associated with the third location to another dimensionality that is greater than or equal to a dimensionality associated with the first vehicle design.
2. The computer-implemented method of claim 1 , wherein generating the first metric comprises computing a distance value between the first location and the second location.
3. The computer-implemented method of claim 1 , wherein the third vehicle design includes at least one design feature associated with the characteristic style.
4. The computer-implemented method of claim 1 , wherein increasing the dimensionality associated with the third location comprises performing a deconvolution operation on the third location.
5. The computer-implemented method of claim 1 , wherein encoding the first vehicle design comprises decreasing the dimensionality associated with the first vehicle design to a dimensionality that is equal to a dimensionality associated with the latent space representation.
6. The computer-implemented method of claim 5 , wherein decreasing the dimensionality associated with the first vehicle design comprises performing a convolution operation on the first vehicle design.
7. The computer-implemented method of claim 1 , wherein the first vehicle design comprises a two-dimensional array of pixels.
8. The computer-implemented method of claim 1 , wherein the first vehicle design comprises a set of non-uniform rational B-spline (NURBS) curves.
9. The computer-implemented method of claim 1 , further comprising performing one or more training operations based on a comparison between different vehicle designs to train the encoder network to project one or more vehicle designs into the latent space representation.
10. The computer-implemented method of claim 1 , further comprising training the encoder network to project designs onto the latent space representation.
11. The computer-implemented method of claim 10 , further comprising training the generator network to generate designs based on the latent space representation.
12. One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of:
projecting a first vehicle design into a latent space representation of vehicle design features to generate a first location;
projecting a second vehicle design into the latent space representation to generate a second location;
determining a distance between the first location and the second location;
generating a style metric based on the distance; and
generating a third vehicle design based on the style metric.
13. The one or more non-transitory computer-readable media of claim 12 , further comprising determining a third location within the latent space representation based on the style metric.
14. The one or more non-transitory computer-readable media of claim 13 , wherein the third vehicle design is generated based on the third location.
15. The one or more non-transitory computer-readable media of claim 14 , wherein the style metric indicates a degree to which the first vehicle design differs from a characteristic style associated with the second vehicle design.
16. The one or more non-transitory computer-readable media of claim 15 , wherein the third vehicle design includes at least one design feature associated with the characteristic style.
17. The one or more non-transitory computer-readable media of claim 16 , wherein generating the third vehicle design comprises:
traversing the latent space representation from the first location to a third location; and
increasing a dimensionality associated with the third location to a dimensionality that is greater than or equal to a dimensionality associated with the first vehicle design.
18. The one or more non-transitory computer-readable media of claim 17 , wherein increasing the dimensionality associated with the third location comprises performing a deconvolution operation on the third location.
19. The one or more non-transitory computer-readable media of claim 18 , projecting the first vehicle design into the latent space representation comprises decreasing a dimensionality associated with the first vehicle design to a dimensionality that is equal to a dimensionality associated with the latent space representation.
20. The one or more non-transitory computer-readable media of claim 19 , wherein decreasing the dimensionality associated with the first vehicle design comprises performing a convolution operation on the first vehicle design.
21. the one or more non-transitory computer readable media of claim 12 , wherein the first vehicle design comprises a two-dimensional array of pixels.
22. The one or more non-transitory computer-readable media of claim 21 , wherein the first vehicle design comprises a set of non-uniform rational B-spline (NURBS) curves.
23. The one or more non-transitory computer-readable media of claim 22 , wherein the first vehicle design incudes at least one sensor feature associated with an autonomous vehicle.
24. A system, comprising:
one or more memories that store instructions; and
one or more processors that are coupled to the one or more memories and,
when executing the instructions, are configured to perform the steps of:
executing an encoder model to project a first vehicle design into a latent space representation of vehicle design features to generate a first location;
executing the encoder model to project a second vehicle design into the latent space representation to generate a second location;
generating a style metric based on the first location and the second location; and
generating a third vehicle design based on the style metric.