Geodesic Convolutional Shape Optimization
Authors: Pierre Baque, Edoardo Remelli, Francois Fleuret, Pascal Fua
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our proposed shape optimization pipeline. It is designed to handle 3D shapes but can also handle 2D ones by simply considering the 2D equivalent of a surface mesh, which is a discretized 2D contour. We therefore first present results on 2D airfoil profiles, which have become a de facto standard in the CFD community for benchmarking shape optimization algorithms (Toal & Keane, 2011; Orman & Durmus, 2016). We then use the example of car shapes to evaluate our algorithm s behavior in the more challenging 3D case. We implemented our deeplearning algorithms in Tensor Flow (Abadi et al., 2016) and ran them on a single Titan X Pascal GPU. |
| Researcher Affiliation | Academia | 1CVLab, EPFL, Lausanne, Switzerland 2Machine Learning Group, Idiap, Martigny, Switzerland. |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the source code for the described methodology or a direct link to a code repository. |
| Open Datasets | No | The paper describes generating its own datasets ('SYNT-TRAIN') and downloading CAD models from the web, but does not provide concrete access information (link, DOI, repository, or formal citation with authors/year) for a publicly available or open dataset. |
| Dataset Splits | No | For 2D shapes, the paper states, 'To generate our training and validation data, we create 8000 training and 8000 testing shapes...', which specifies the size of training and testing sets, but does not provide a specific, distinct split (e.g., percentage or count) for validation data separate from the training set. For 3D shapes, datasets are described (SYNT-TRAIN, SYNT-TEST, CARS-Fine Tune, CARS-TEST) but without explicit train/validation/test percentages or counts for splitting. |
| Hardware Specification | Yes | We implemented our deeplearning algorithms in Tensor Flow (Abadi et al., 2016) and ran them on a single Titan X Pascal GPU. |
| Software Dependencies | Yes | We implemented our deeplearning algorithms in Tensor Flow (Abadi et al., 2016) and ran them on a single Titan X Pascal GPU. [...] We use the popular CFD simulator XFoil to compute their aerodynamic properties. [...] We used the industry standard Ansys Fluent (Inc., 2011) to compute their aerodynamic properties with the k-epsilon turbulence model. |
| Experiment Setup | Yes | Let K, where K = 32 in all our experiments, be a predefined number of gaussian parameters αk R2, Σk R2, which are vertex independent. [...] In our experiments, we set λ = 100 and Ztarget = 0.8. [...] For KNN regression, we empirically determined that K = 8 combined to a distance-based neighbor weighting yielded the best results. |