Differentiable Spline Approximations

Authors: Minsu Cho, Aditya Balu, Ameya Joshi, Anjana Deva Prasad, Biswajit Khara, Soumik Sarkar, Baskar Ganapathysubramanian, Adarsh Krishnamurthy, Chinmay Hegde

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show applications of our approach in three stylized applications: image segmentation, 3D point cloud reconstruction, and finite element analysis for the solution of partial differential equations. 3 Experiments We have implemented the DSA framework (and its different applications provided below)...
Researcher Affiliation Academia Minsu Cho1 Aditya Balu2 Ameya Joshi1 Anjana Deva Prasad2 Biswajit Khara2 Soumik Sarkar2 Baskar Ganapathysubramanian2 Adarsh Krishnamurthy2 Chinmay Hegde1 New York University1, Iowa State University2 {mc8065, ameya.joshi, chinmay.h}@nyu.edu {baditya, anjana, bkhara, soumiks, baskarg, adarsh}@iastate.edu
Pseudocode Yes Algorithm 1 Backward pass for NURBS Jacobian (for one curve point , C(u))
Open Source Code Yes We also open-source the code at https://github.com/idealab-isu/DSA.
Open Datasets Yes We train two models (MDSA, Mbaseline) on two different segmentation tasks: the Weizmann horse dataset [Borenstein and Ullman, 2004] and the Broad Bioimage Benchmark Collection dataset [Ljosa et al., 2012] (publicly available under Creative Commons License). The Spline Dataset, which is a subset of surfaces extracted from the ABC dataset (available for public use under this license).
Dataset Splits No The paper mentions splitting datasets into train (85%) and test (15%) for the Weizmann horse and Broad Bioimage Benchmark Collection datasets, but does not explicitly state a separate validation split or its details.
Hardware Specification Yes All the experiments were performed using a local cluster with 6 compute nodes and each node having 2 GPUs (Tesla V100s with 32GB GPU memory).
Software Dependencies No The paper mentions 'extending autograd functions in Pytorch' and 'CUDA for GPU support', but does not specify version numbers for PyTorch, CUDA, or other key software components.
Experiment Setup Yes We use the same architecture and hyper-parameters for both models (see Appendix for details.) All training was done using a single GPU.